INTERNET-DRAFT John Lazzaro January 15, 2003 John Wawrzynek Expires: July 15, 2003 UC Berkeley An Implementation Guide to the MIDI Wire Protocol Packetization (MWPP) Status of this Memo This document is an Internet-Draft and is subject to all provisions of Section 10 of RFC2026. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/1id-abstracts.html The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html Abstract This memo offers non-normative implementation guidance for the MIDI Wire Protocol Packetization (MWPP), an RTP packetization for the MIDI command language. In the main body of the memo, we discuss one MWPP application in detail: an interactive, two-party, single- stream session over unicast UDP transport that uses RTCP. In the Appendices, we discuss specialized implementation issues: MWPP without RTCP, MWPP with TCP, multi-stream sessions, multi-party sessions, and content streaming. Lazzaro/Wawrzynek [Page 1] INTERNET-DRAFT 15 January 2003 0. Change Log for This version is the first complete release of the document. Most sections of -00.txt have been rewritten. Thanks to Phil Kerr for suggestions on improving the two-party sender description (Sections 4 and 5 of the main text), and to Joanne Dow for suggestions on explaining timing issues. Sections that are new to this document: o Sections 6 and 7 describe MWPP receiver algorithms for a two-party unicast interactive session. o Appendix A describes content streaming, including the use of RTSP, and the use of an ancillary FEC or retransmission stream. o Appendix B describes multi-party interactive sessions, and discusses two transport options: unicast meshes and multicast groups. o Appendix C describes TCP and MWPP. It shows how to set up TCP sessions using the comedia mmusic I-D. It also covers RTSP interleave mode. o Appendix D describes how to use MWPP without RTCP over UDP. o Appendix E describes sessions that use several MWPP streams, and shows synchronization and MIDI name space techniques. Examples show multi-stream sessions to implement common applications, such as session archiving. The next step is a comprehensive pass over the normative MWPP document to address remaining open issues. I'll then do another pass over this informative I-D, to realign it with the normative text, and to add in any suggestions I get from the WG and others. As always, feedback is much appreciated, on the list or to lazzaro@cs.berkeley.edu. Lazzaro/Wawrzynek [Page 2] INTERNET-DRAFT 15 January 2003 Table of Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2. Session Management: Starting MWPP Sessions . . . . . . . . . . . 6 3. Session Management: Session Housekeeping . . . . . . . . . . . . 12 4. Sending MWPP Streams: General Considerations . . . . . . . . . . 13 4.1 Queuing and Coding Incoming MIDI Data . . . . . . . . . . . 14 4.2 Sending MWPP Packets with Empty MIDI Lists . . . . . . . . 15 4.3 Bandwidth Management and Congestion Control . . . . . . . . 16 5. Sending MWPP Streams: The Recovery Journal . . . . . . . . . . . 18 5.1 Initializing the RJSS . . . . . . . . . . . . . . . . . . . 21 5.2 Traversing the RJSS . . . . . . . . . . . . . . . . . . . . 21 5.3 Updating the RJSS . . . . . . . . . . . . . . . . . . . . . 22 5.4 Trimming the RJSS . . . . . . . . . . . . . . . . . . . . . 23 5.5 Implementation Notes . . . . . . . . . . . . . . . . . . . 24 6. Receiving MWPP Streams: General Considerations . . . . . . . . . 25 6.1 The NMP Receiver Design . . . . . . . . . . . . . . . . . . 26 6.2 Receiver Design Issues . . . . . . . . . . . . . . . . . . 28 7. Receiving MWPP Streams: The Recovery Journal . . . . . . . . . . 29 7.1 Chapter W: MIDI Pitch Wheel (0xE) . . . . . . . . . . . . . 32 7.2 Chapter N: MIDI NoteOn (0x8) and NoteOff (0x9) . . . . . . 33 7.3 Chapter C: MIDI Control Change (0xB) . . . . . . . . . . . 35 7.4 Chapter P: MIDI Program Change (0xC) . . . . . . . . . . . 36 8. Congestion Control . . . . . . . . . . . . . . . . . . . . . . . 38 9. Security Considerations . . . . . . . . . . . . . . . . . . . . . 38 10. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . 38 Appendix A. Content Streaming with MWPP . . . . . . . . . . . . . . 39 A.1 Session Management . . . . . . . . . . . . . . . . . . . . 39 A.2 Baseline Algorithms . . . . . . . . . . . . . . . . . . . . 41 A.3 Packet Replacement Streams . . . . . . . . . . . . . . . . 43 Appendix B. Multi-party MWPP Sessions . . . . . . . . . . . . . . . 45 B.1 Session Management (simulated multicast) . . . . . . . . . 45 B.2 Session Management (true multicast) . . . . . . . . . . . . 48 B.3 Sender Issues . . . . . . . . . . . . . . . . . . . . . . . 49 B.4 Receiver Issues . . . . . . . . . . . . . . . . . . . . . . 51 B.5 Scaling Issues . . . . . . . . . . . . . . . . . . . . . . 52 Appendix C. MWPP and Reliable Transport . . . . . . . . . . . . . . 54 C.1 Session Management . . . . . . . . . . . . . . . . . . . . 55 C.2 Sending and Receiving . . . . . . . . . . . . . . . . . . . 57 C.3 RTSP Interleaving . . . . . . . . . . . . . . . . . . . . . 57 Appendix D. Using MWPP without RTCP . . . . . . . . . . . . . . . . 59 D.1 Session Management . . . . . . . . . . . . . . . . . . . . 59 D.2 Sender Issues . . . . . . . . . . . . . . . . . . . . . . . 61 D.3 Receiver Issues . . . . . . . . . . . . . . . . . . . . . . 62 Appendix E. Multi-stream MWPP Sessions . . . . . . . . . . . . . . . 64 E.1 Session Scenarios . . . . . . . . . . . . . . . . . . . . . 64 E.2 Synchronization Issues . . . . . . . . . . . . . . . . . . 71 Lazzaro/Wawrzynek [Page 3] INTERNET-DRAFT 15 January 2003 E.3 Name Space Issues . . . . . . . . . . . . . . . . . . . . . 73 Appendix F. References . . . . . . . . . . . . . . . . . . . . . . . 75 F.1 Normative References . . . . . . . . . . . . . . . . . . . 75 F.2 Informative References . . . . . . . . . . . . . . . . . . 76 Appendix G. Author Addresses . . . . . . . . . . . . . . . . . . . . 77 Lazzaro/Wawrzynek [Page 4] INTERNET-DRAFT 15 January 2003 1. Introduction The MIDI Wire Protocol Packetization (MWPP, [1]) is a general-purpose RTP/AVP [2,3] packetization for the MIDI [4] command language. [1] normatively defines the MWPP RTP bitfield syntax, and also defines the Session Description Protocol (SDP, [5]) parameters that may be used to customize MWPP session behavior. However, [1] does not define algorithms for sending and receiving MWPP streams. Implementors are free to use any sending or receiving algorithm that conforms to the normative text in [1]. In this memo, we offer advice on how to implement sending, receiving, and session management algorithms for MWPP. Unlike [1], this memo is not normative. The application space for MWPP is diverse, and may be categorized in the following ways: o Interactive or streaming. Interactive applications (such as the remote operation of musical instruments) require low end-to-end latency, preferably near the underlying network latency. Streaming applications (such as the incremental delivery of MIDI files) trade off higher latency for better fidelity and efficiency. o Two-party or multi-party. Two-party MWPP applications have two session participants; multi-party MWPP applications have more than two participants. Multi-party applications map efficiently to multicast transport, but may also use multiple unicast flows. o Transport. MWPP streams may use unreliable transport (such as unicast or multicast UDP) or reliable transport (such as TCP). o Single-stream or multi-stream. Simple MWPP sessions use one RTP stream to convey a single MIDI name space (16 voice channels + systems). Multi-stream sessions use several RTP streams to convey more than 16 voice channels. Multi-stream sessions are also used to split a MIDI name space across different transport types. o RTCP or no RTCP. The RTP standard [2] defines a backchannel protocol, the RTP Control Protocol (RTCP). MWPP RTP streams work best if paired with an RTCP stream, but MWPP may be used without RTCP. Lazzaro/Wawrzynek [Page 5] INTERNET-DRAFT 15 January 2003 In the main body of this memo, we describe an interactive, two-party, single-stream session over unicast UDP transport that uses RTCP. Sections 2 and 3 cover session management; Sections 4 and 5 cover sending MWPP streams; Sections 6 and 7 cover receiving MWPP streams. The main text is written with a specific application in mind: network musical performance over wide-area networks. As defined in [13], a network musical performance occurs when a group of musicians, located at different physical locations, interact over a network to perform as they would if located in the same room. However, the methods we describe in the main text are also applicable to local-area network (LAN) applications, such as the remote control of musical instruments. The main text includes several discussions of LAN issues, such as LAN receiver design guidance in Section 6. In the Appendices of this memo, we discuss implementation issues for other session types. For example, Appendix A describes implementation issues in content streaming. Each Appendix covers session management, sender design, and receiver design. This memo is limited in scope, in that it assumes that all session participants have access to the SDP session description(s) that describe the session. We do not discuss the creation, negotiation, or distribution of session descriptions, apart from a discussion of the Real Time Streaming Protocol (RTSP, [6]) in Appendix A. We anticipate that other memos will define frameworks for session description issues for MWPP, and that these memos will include implementation guidance. 2. Session Management: Starting MWPP Sessions In this section, we discuss how interactive MWPP applications start sessions. We limit our discussion to two-party sessions over unicast UDP transport that use RTCP. In the Appendices, we discuss startup issues for other types of sessions. We assume that the two parties have agreed on a session configuration, embodied by a pair of Session Description Protocol (SDP, [5]) session descriptions. One session description (Figure 1) defines how the first party wishes to receive its stream; the other session description (Figure 2) defines how the second party wishes to receive its stream. Even if one party is not receiving an RTP stream (indicated by the SDP attribute sendonly [5]), the party still defines a session description, in order to describe how it receives its RTCP stream. Lazzaro/Wawrzynek [Page 6] INTERNET-DRAFT 15 January 2003 v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=Example t=0 0 c=IN IP4 192.0.2.94 m=audio 16112 RTP/AVP 96 a=rtpmap: 96 mwpp/44100 Figure 1 -- Session description for first participant. v=0 o=second 2520644554 2838152170 IN IP4 second.example.net s=Example t=0 0 c=IN IP4 192.0.2.105 m=audio 5004 RTP/AVP 101 a=rtpmap: 101 mwpp/44100 Figure 2 -- Session description for second participant. The session description in Figure 1 codes that the first party intends to receive an MWPP RTP stream on IP4 number 192.0.2.94 (coded in the c= line) at UDP port 16112 (coded in the m= line). Implicit in the SDP m= line syntax [5] is that the first party also intends to receive an RTCP stream on 192.0.2.94 at UDP port 16113 (16112 + 1). The receiver expects that the PTYPE field of each RTP header in the received stream will be set to 96 (coded in the m= and a= lines). Likewise, the session description in Figure 2 codes that the second party intends to receive an MWPP RTP stream on IP4 number 192.0.2.105 at UDP port 5004, and also intends to receive an RTCP stream on 192.0.2.105 at UDP port 5005 (5004 + 1). The second party expects that the PTYPE RTP header field of received stream will be set to 101. The session descriptions do not use the SDP parameter render (Appendix A.5 of [1]) to indicate the rendering method for the MIDI stream. If render was in use, the parties would use this information to set up the appropriate rendering algorithms for the MIDI stream. We now show example code that implements the actions the parties take during the session. The code is written in C, and uses the sockets API and other POSIX systems calls. We show code for the first party (the second party takes a symmetric set of actions). Lazzaro/Wawrzynek [Page 7] INTERNET-DRAFT 15 January 2003 Figure 3 shows how the first party initializes a pair of socket descriptors (rtp_fd and rtcp_fd) to send and receive UDP packets. The code sets up the descriptors to listen to ports 16112 and 16113 on the IP4 network connection for 192.0.2.94. Note that the code assumes a single-homed machine. The ERROR_RETURN macro is used to flag fatal setup errors (this macro is not defined in Figure 3). After the code in Figure 3 runs, the first party may check for new RTP or RTCP packets by calling recv() on rtp_fd or rtcp_fd. By default, a recv() call on these socket descriptors blocks until a packet arrives. Figure 4 shows how configure these sockets as non-blocking, so that recv() calls may be done in time-critical code without fear of I/O blocking. Figure 5 shows how to use recv() to check a non-blocking socket for new packets. The first party also uses rtp_fd and rtcp_fd to send RTP and RTCP packets to the second party. In Figure 6, we show how to initialize socket structures that address the second party. In Figure 7, we show how to use one of these structures in a sendto() call to send an RTP packet to the second party. Note that the code shown in Figures 3-7 assumes a clear network path between the participants. The code may not work if firewalls or Network Address Translation (NAT) devices are present in the network path. See [15] for standardized methods for overcoming network obstacles. Lazzaro/Wawrzynek [Page 8] INTERNET-DRAFT 15 January 2003 #include #include #include int rtp_fd, rtcp_fd; /* socket descriptors */ struct sockaddr_in addr; /* for bind address */ /*********************************/ /* create the socket descriptors */ /*********************************/ if ((rtp_fd = socket(AF_INET, SOCK_DGRAM, 0)) < 0) ERROR_RETURN("Couldn't create Internet RTP socket"); if ((rtcp_fd = socket(AF_INET, SOCK_DGRAM, 0)) < 0) ERROR_RETURN("Couldn't create Internet RTCP socket"); /**********************************/ /* bind the RTP socket descriptor */ /**********************************/ memset(&(addr.sin_zero), 0, 8); addr.sin_family = AF_INET; addr.sin_addr.s_addr = htonl(INADDR_ANY); addr.sin_port = htons(16112); /* port 16112, from SDP */ if (bind(rtp_fd, (struct sockaddr *)&addr, sizeof(struct sockaddr)) < 0) ERROR_RETURN("Couldn't bind Internet RTP socket"); /***********************************/ /* bind the RTCP socket descriptor */ /***********************************/ memset(&(addr.sin_zero), 0, 8); addr.sin_family = AF_INET; addr.sin_addr.s_addr = htonl(INADDR_ANY); addr.sin_port = htons(16113); /* port 16113, from SDP */ if (bind(rtcp_fd, (struct sockaddr *)&addr, sizeof(struct sockaddr)) < 0) ERROR_RETURN("Couldn't bind Internet RTCP socket"); Figure 3 -- Setup code for listening for RTP/RTCP packets. Lazzaro/Wawrzynek [Page 9] INTERNET-DRAFT 15 January 2003 #include #include int one = 1; /*******************************************************/ /* set non-blocking status, shield spurious ICMP errno */ /*******************************************************/ if (fcntl(rtp_fd, F_SETFL, O_NONBLOCK)) ERROR_RETURN("Couldn't unblock Internet RTP socket"); if (fcntl(rtcp_fd, F_SETFL, O_NONBLOCK)) ERROR_RETURN("Couldn't unblock Internet RTCP socket"); if (setsockopt(rtp_fd, SOL_SOCKET, SO_BSDCOMPAT, &one, sizeof(one))) ERROR_RETURN("Couldn't shield RTP socket"); if (setsockopt(rtcp_fd, SOL_SOCKET, SO_BSDCOMPAT, &one, sizeof(one))) ERROR_RETURN("Couldn't shield RTCP socket"); Figure 4 -- Code to set socket descriptors to be non-blocking. #include #define UDPMAXSIZE 1472 /* based on Ethernet MTU of 1500 */ unsigned char packet[UDPMAXSIZE+1]; int len; while ((len = recv(rtp_fd, packet, UDPMAXSIZE + 1, 0)) > 0) { /* process packet[], be cautious if (len == UDPMAXSIZE + 1) */ } if ((len == 0) || (errno != EAGAIN)) { /* while() may have exited in an unexpected way */ } Figure 5 -- Code to check rtp_fd for new RTP packets. Lazzaro/Wawrzynek [Page 10] INTERNET-DRAFT 15 January 2003 #include #include struct sockaddr_in * rtp_addr; /* RTP destination IP/port */ struct sockaddr_in * rtcp_addr; /* RTCP destination IP/port */ /* set RTP address, as coded in Figure 2's SDP */ rtp_addr = calloc(1, sizeof(struct sockaddr_in)); rtp_addr->sin_family = AF_INET; rtp_addr->sin_port = htons(5004); rtp_addr->sin_addr.s_addr = inet_addr("192.0.2.105"); /* set RTCP address, as coded in Figure 2's SDP */ rtcp_addr = calloc(1, sizeof(struct sockaddr_in)); rtcp_addr->sin_family = AF_INET; rtcp_addr->sin_port = htons(5005); /* 5004 + 1 */ rtcp_addr->sin_addr.s_addr = rtp_addr->sin_addr.s_addr; Figure 6 -- Initializing destination addresses for RTP and RTCP. unsigned char packet[UDPMAXSIZE]; /* RTP packet to send */ int size; /* length of RTP packet */ /* first fill packet[] and set size ... then: */ if (sendto(rtp_fd, packet, size, 0, rtp_addr, sizeof(struct sockaddr)) == -1) { /* * try again later if errno == EAGAIN or EINTR * * other errno values --> an operational error */ } Figure 7 -- Using sendto() to send an RTP packet. Lazzaro/Wawrzynek [Page 11] INTERNET-DRAFT 15 January 2003 3. Session Management: Session Housekeeping After the two-party interactive session is set up, the parties begin to send and receive MWPP RTP packets. In Sections 4-7, we discuss MWPP RTP sending and receiving algorithms. In this section, we describe session "housekeeping" tasks that the participants also perform. One housekeeping function is the maintenance of the 32-bit SSRC value that uniquely identifies each party. Section 8 of [2] describes SSRC issues in detail. Another housekeeping function is the sending and receiving of RTCP. MWPP uses the standard techniques for sending and receiving RTCP, which are described in Section 6 of [2]. However, MWPP defines the sampling instant of an RTP packet in an unusual way (Section 2.1 of [1]), affecting the calculation of RTCP reception statistics. Another housekeeping function concerns security. As detailed in the Security Considerations section of [1], per-packet authentication is strongly recommended for use with MWPP, because the acceptance of rogue MWPP packets may lead to the execution of arbitrary MIDI commands. [16] describes a standard for authenticating RTP and RTCP packets. To simplify the presentation of sending and receiving algorithms in this memo, our examples do not authenticate packets. A final housekeeping function concerns the termination of an MWPP RTP session. In our two-party example, the session terminates upon the exit of one of the participants. A clean termination may require active effort by a receiver, as a MIDI stream stopped at an arbitrary point may cause stuck notes and other indefinite artifacts in the MIDI renderer. The exit of a party may be signalled in several ways. Session management tools may offer a reliable signal for termination (such as the SIP BYE method [14]). The (unreliable) RTCP BYE packet [2] may also signal the exit of a party. Receivers may also sense the lack of RTCP activity and timeout a party, or may use transport methods to detect an exit. Lazzaro/Wawrzynek [Page 12] INTERNET-DRAFT 15 January 2003 4. Sending MWPP Streams: General Considerations In this section we discuss sender implementation issues, for a two-party interactive session. The session represents a network musical performance between two players over a wide-area network. An interactive MWPP sender is a real-time data-driven entity. On an on- going basis, the sender checks to see if the local player has generated new MIDI data. At any time, the sender may transmit a new MWPP RTP packet to the remote player, for the reasons described below: 1. New MIDI data has been generated by the local player, and the sender decides it is time to issue a packet coding the data. 2. The local player has not generated new MIDI data, but the sender decides too much time has elapsed since the last RTP packet transmission. The sender transmits a packet in order to relay updated header and recovery journal data. In both cases, the sender generates a packet that consists of an RTP header, a MIDI Command section, and a recovery journal. In the first case, the MIDI list of the MIDI Command section codes the new MIDI data. In the second case, the MIDI list is empty. Figure 8 shows the 5 steps a sender takes to issue a packet. Algorithm for Sending an MWPP Packet: 1. Generate the RTP header for the new packet. See Section 2.1 of [1] for details. 2. Generate the MIDI Command section for the new packet. See Section 3 of [1] for details. 3. Generate the recovery journal for the new packet. We discuss this process in Section 5.2. The generation algorithm examines the Recovery Journal Sending Structure (RJSS), a stateful coding of a history of the stream. 4. Send the new packet to the receiver. 5. Update the RJSS to include the data coded in the MIDI Command section of the packet sent in step 4. We discuss the update procedure in Section 5.3. Figure 8 -- A 5 step algorithm for sending an MWPP RTP packet. Lazzaro/Wawrzynek [Page 13] INTERNET-DRAFT 15 January 2003 The algorithm shown in Figure 8 corresponds to the code fragment for sending RTP packets shown in Figure 7 of Section 2. Steps 1, 2, and 3 occur before the sendto() call in the code fragment. Step 4 corresponds to the sendto() call itself. Step 5 may occur once Step 3 completes. In the sections that follow, we discuss specific sender implementation issues in detail. 4.1 Queuing and Coding Incoming MIDI Data In this section, we describe how a sender decides when to transmit a new RTP packet. We also discuss sender timestamp coding issues. Simple senders transmit a new MWPP RTP packet as soon as the local player generates a complete MIDI command. The system described in [13] uses this algorithm. This algorithm has zero sender queuing latency, as the sender never delays the transmission of a new MIDI command. In a relative sense, this algorithm uses bandwidth inefficiently, as it does not amortize the overhead of an MWPP RTP packet over several MIDI commands. This inefficiency may be acceptable for sparse MIDI data streams (see Appendix A.4 of [13]). More sophisticated sending algorithms [17] improve efficiency by coding small groups of MIDI commands into a single RTP packet, at the expense of non-zero sender queuing latency. Senders assign a timestamp value to each MIDI command in the stream. The default timestamp semantics are defined in Section 3 of [1]. The SDP parameters tsmode, linerate, octpos, and mperiod (Appendix C.2 of [1]) may be used to customize timestamp semantics during session setup. Senders may code the timestamp values for MIDI commands in two ways. The most efficient method is to set the RTP timestamp of the packet to the timestamp of the first command in the MIDI list. In this method, the Z bit of the MIDI command section header (Figure 2 of [1]) is set to 0. The RTP timestamps of the stream increment at a non-uniform rate. In some applications, senders may wish to generate a stream whose RTP timestamps increment at a uniform rate (perhaps to improve the performance of header compression [18]). To code the timestamp of the first command in the MIDI list, the sender uses the optional delta time field. The Z bit of the MIDI command section header is set to 1. Finally, as we discuss in Section 6, interactive receivers may model the network latency, and use the model to optimize its rendering performance. By necessity, models use the timestamp of the last command coded in the MIDI list as a proxy for the sending time. If the MIDI list is empty, the RTP timestamp serves as the proxy. Lazzaro/Wawrzynek [Page 14] INTERNET-DRAFT 15 January 2003 To the extent possible, interactive senders should maintain a constant relationship between this proxy and the actual sending time. To the receiver, variance in this relationship is indistinguishable from network jitter. 4.2 Sending MWPP Packets with Empty MIDI Lists As we described in the preamble of Section 4, interactive senders may decide to transmit MWPP RTP packets with empty MIDI lists. Senders generate "empty packets" in two contexts: as "keep-alive" packets during periods of no MIDI activity, and as "guard" packets to improve the performance of the recovery journal system. In this section, we discuss implementation issues for empty packets. In an interactive session, musicians might refrain from generating MIDI data for extended periods of time (seconds or even minutes). If an MWPP RTP stream followed the dynamics of a silent MIDI source, and stopped sending RTP packets for an extended periods, systems behavior might be degraded in the following ways: o Receivers may misinterpret the silent stream as a dropped network connection. o Network middleboxes (such as Network Address Translators) may "time-out" the silent stream and drop the port and IP association state. o The receiver's model of network performance may fall out of date. Senders avoid these problems by sending "keep-alive" MWPP packets during periods of network inactivity. Keep-alive packets have empty MIDI lists. Session participants may specify the frequency of keep-alive packets during session configuration with the SDP parameter maxptime (Appendix C.3 of [1]). As a point of reference, the system described in [13] sends a keep-alive packet if no RTP packet has been sent for 30 seconds. Senders may also send empty MWPP packets to improve the performance of the recovery journal system. As we describe in Section 6, the recovery process begins when a receiver detects a break in the RTP sequence number pattern of the stream. The receiver uses the recovery journal of the break packet to guide corrective rendering actions, such as ending stuck notes and updating out-of-date controller values. Lazzaro/Wawrzynek [Page 15] INTERNET-DRAFT 15 January 2003 Consider the situation where the local player produces a MIDI NoteOff command (which the sender promptly transmits in an MWPP packet), but then 5 seconds pass before the player produces another MIDI command (which the sender transmits in a second MWPP packet). If the MWPP packet coding the NoteOff is lost, the receiver will not be aware of the packet loss incident for 5 seconds, and the rendered MIDI performance will contain a note that sounds for 5 seconds too long. To handle this situation, senders may transmit empty MWPP packets to "guard" the stream during silent sections. The guard packet algorithm defined in Section 7.3 of [13], as applied to the situation described above, would send a guard packet after 100 ms of player inactivity, and would send a second guard packet 100 ms. later. Subsequent guard packets would be sent with an exponential backoff, with a limiting period of 1 second. Guard packet transmissions would cease once MIDI activity resumes, or once RTCP receiver reports indicate that the receiver is up to date. We view the perceptual quality of guard packet sending algorithms as a quality of implementation factor for MWPP applications. Sophisticated implementations may tailor the guard packet sending rate to the nature of the MIDI commands recently sent in the stream, to minimize the perceptual impact of moderate packet loss. As an example of this sort of specialization, the guard packet algorithm described in [13] protects against the transient artifacts that occur when NoteOn MIDI commands are lost. The algorithm sends a guard packet 1 ms after an MWPP packet whose MIDI list contains a NoteOn command. The Y bit in Chapter N note logs (Appendix A.4 of [1]) supports this use of guard packets. Bandwidth management and congestion control are key issues in guard packet algorithms. We discuss these issues in the next section. 4.3 Bandwidth Management and Congestion Control Senders may control the instantaneous sending rate of an MWPP stream in a variety of ways. In this section, we describe the mechanics of MWPP rate control, in the contexts of congestion control and bandwidth management. RTP implementations have a responsibility to implement congestion control mechanisms to protect the network infrastructure (see Section 10 of [2]). In general, senders implement congestion control by monitoring packet loss via RTCP receiver reports, and reducing the stream sending rate if packet loss is excessive. Section 6.4.4 of [2] provides guidance for using the RTCP receiver report fields for congestion control. Lazzaro/Wawrzynek [Page 16] INTERNET-DRAFT 15 January 2003 Bandwidth management is a second use for MWPP sending rate control. An SDP session description may optionally include a bandwidth line (b=, as defined in Section 6 of [5]) to specify the maximum bandwidth an RTP stream may use. If an MWPP session description includes a bandwidth line, senders control the instantaneous sending rate of the stream so that the maximum bandwidth is not exceeded. Interactive MWPP senders have a variety of methods to control the instantaneous sending rate: o As described in Section 4.1, MWPP senders may pack several MIDI commands into a single MWPP packet, thereby reducing instantaneous stream bandwidth at the expense of increasing sender queuing latency. o Guard packet algorithms (Section 4.2) may be designed in a parametric way, so that the tradeoff between artifact reduction and stream bandwidth may be tuned dynamically. o The recovery journal size may be reduced, by adapting the techniques described in Section 5 of this memo and in Section 4.1 of [1]. Note that in all cases, the recovery journal sender must conform to the mandate defined in Section 4 of [1]. o The incoming MIDI stream may be modified, to reduce the number of MIDI commands without significantly altering the MIDI performance. Lossy "MIDI filtering" algorithms are well developed in the MIDI community, and may be directly applied to MWPP rate management. MWPP senders incorporate these rate control methods into feedback loops to implement congestion control and bandwidth management. Lazzaro/Wawrzynek [Page 17] INTERNET-DRAFT 15 January 2003 5. Sending MWPP Streams: The Recovery Journal In this section, we describe how senders implement the recovery journal system. We begin by describing the Recovery Journal Sending Structure (RJSS). Senders use the RJSS to generate the recovery journal section for MWPP RTP packets. The RJSS is a hierarchical representation of the checkpoint history of the stream. The checkpoint history holds the MIDI commands that are at risk to packet loss (see Appendix A.1 of [1] for a precise definition of the checkpoint history). The layout of the RJSS mirrors the hierarchical structure of the recovery journal bitfields. Figure 9 shows a RJSS implementation for a simple MWPP sender. The sender transmits most voice command types, but does not transmit system commands. The leaf level of the hierarchy (the jsend_chapter structures) corresponds to channel chapters (Appendices A.2-7 in [1]). The second level of the hierarchy (jsend_channel) corresponds to the channel journal header (Figure 8 in [1]). The top level of the hierarchy (jsend_journal) corresponds to the recovery journal header (Figure 7 in [1]). Each level in the RJSS may code several items: 1. The current contents of the recovery journal bitfield for the level (jheader[], cheader[], and the chapter bitfields). 2. A seqnum variable. Seqnum codes the extended RTP sequence number of the most recent packet that added information to the checkpoint history, at the level or at any level below it. A seqnum variable is set to zero if the checkpoint history contains no information at the level or at any level below it. 3. Ancillary variables used by the sending algorithm. In the sections that follow, we describe the tasks a sender performs to manage the recovery journal system. Lazzaro/Wawrzynek [Page 18] INTERNET-DRAFT 15 January 2003 typedef unsigned char uint8; /* must be 1 octet */ typedef unsigned short uint16; /* must be 2 octet */ typedef unsigned long uint32; /* must be 4 octets */ /***********************************************************/ /* leaf level of hierarchy: Chapter W, Appendix A.3 of [1] */ /***********************************************************/ typedef struct jsend_chapterw { /* Pitch Wheel (0xE) */ uint8 chapterw[2]; /* bitfield (Figure A.3.1, [1]) */ uint32 seqnum; /* extended sequence number, or 0 */ } jsend_chapterw; /***********************************************************/ /* leaf level of hierarchy: Chapter N, Appendix A.4 of [1] */ /***********************************************************/ typedef struct jsend_chaptern { /* Note commands (0x8, 0x9) */ uint8 chaptern[272]; /* bitfield (Figure A.4.1, [1]) */ uint16 size; /* actual size of chaptern[] */ uint32 seqnum; /* extended sequence number, or 0 */ uint32 note_seqnum[128]; /* most recent note seqnum, or 0 */ uint32 note_tstamp[128]; /* NoteOn execution timestamp */ uint8 note_state[128]; /* NoteOn velocity, 0 -> NoteOff */ } jsend_chaptern; /***********************************************************/ /* leaf level of hierarchy: Chapter C, Appendix A.7 of [1] */ /***********************************************************/ typedef struct jsend_chapterc { /* Control Change (0xB) */ uint8 chapterc[257]; /* bitfield (Figure A.7.1, [1]) */ uint16 size; /* actual size of chapterc[] */ uint32 seqnum; /* extended sequence number, or 0 */ uint8 control_state[128]; /* per-number control state */ uint32 control_seqnum[128]; /* most recent seqnum, or 0 */ } jsend_chapterc; Figure 9 -- Recovery Journal Sending Structure (part 1) Lazzaro/Wawrzynek [Page 19] INTERNET-DRAFT 15 January 2003 /***********************************************************/ /* leaf level of hierarchy: Chapter P, Appendix A.2 of [1] */ /***********************************************************/ typedef struct jsend_chapterp { /* MIDI Program Change (0xC) */ uint8 chapterp[3]; /* bitfield (Figure A.2.1, [1]) */ uint32 seqnum; /* extended sequence number, or 0 */ } jsend_chapterp; /***************************************************/ /* second-level of hierarchy, for channel journals */ /***************************************************/ typedef struct jsend_channel { uint8 cheader[3]; /* header bitfield (Figure 8, [1]) */ uint32 seqnum; /* extended sequence number, or 0 */ jsend_chapterp chapterp; /* chapter P info */ jsend_chapterw chapterw; /* chapter W info */ jsend_chaptern chaptern; /* chapter N info */ jsend_chapterc chapterc; /* chapter C info */ } jsend_channel; /*******************************************************/ /* top level of hierarchy, for recovery journal header */ /*******************************************************/ typedef struct jsend_journal { uint8 jheader[3]; /* header bitfield (Figure 7, [1]) */ /* Note: Empty journal has a header */ uint32 seqnum; /* extended sequence number, or 0 */ /* seqnum = 0 codes empty journal */ jsend_channel channels[16]; /* channel journal state */ /* index is MIDI channel */ } jsend_journal; Figure 9 (continued) -- Recovery Journal Sending Structure Lazzaro/Wawrzynek [Page 20] INTERNET-DRAFT 15 January 2003 5.1 Initializing the RJSS At the start of a stream, the sender initializes the RJSS. All seqnum variables are set to zero, including all elements of note_seqnum[] and control_seqnum[]. The sender initializes jheader[] to form a recovery journal header that codes an empty journal. The S bit of the header is set to 1, and the A, Y, R, and TOTCHAN header fields are set to zero. The checkpoint packet sequence number field is set to the sequence number of the upcoming first RTP packet (per Appendix A.1 of [1]). In jsend_chaptern, elements of note_tstamp[] and note_state[] are set to zero. In jsend_chapterc, control_state[] is initialized to the default value for each controller number, in the format of the chosen tool type (as defined in Appendix A.7 in [1]). 5.2 Traversing the RJSS Whenever an MWPP RTP packet is created (Step 3 in the algorithm defined in Figure 8), the sender traverses the RJSS to create the recovery journal for the packet. The traversal begins at the top level of the RJSS. The sender copies jheader[] into the packet, and then sets the S bit of jheader[] to 1. The traversal continues depth-first, visiting every jsend_channel whose seqnum variable is non-zero. The sender copies the cheader[] array into the packet, and then sets the S bit of cheader[] to 1. After each cheader[] copy, the sender visits each leaf-level chapter, in order of its appearance in the chapter journal Table of Contents (first P, then W, then N, then C, as shown in Figure 8 of [1]). If a chapter has a non-zero seqnum, the sender copies the chapter bitfield array into the packet, and then sets the S bit of the RJSS array to 1. For chaptern[], the B bit is also set to 1. For the variable-length chapters (chaptern[] and chapterc[]), the sender checks the size variable to determine the bitfield length Before copying chaptern[], the sender updates the Y bit of each note log to code the onset of the associated NoteOn command (Figure A.4.3 in [1]). To determine the Y bit value, the sender checks the note_tstamp[] array for note timing information. Lazzaro/Wawrzynek [Page 21] INTERNET-DRAFT 15 January 2003 5.3 Updating the RJSS After an MWPP RTP packet is sent, the sender updates the RJSS to refresh the checkpoint history (Step 5 in the sending algorithm defined in Figure 8). For each command in the MIDI list of the sent packet, the sender performs the update procedure we describe below. The update procedure begins at the leaf level. The sender generates a new bitfield array for the chapter associated with the MIDI command, using the chapter-specific semantics defined in Appendix A of [1]. For the fixed-length chapterp[] or chapterw[], the sender operates directly on the bitfields. For the variable-length chaptern[] or chapterc[], the sender uses a two-step update algorithm: 1. The sender updates the state arrays for the command note number (Chapter N) or controller number (Chapter C). These arrays, in jsend_chaptern or jsend_chapterc in Figure 9, code the packet extended sequence number (note_seqnum[] and control_seqnum[]), the command execution timestamp (note_tstamp[]), and information from the command data field (note_state[] or control_state[]). 2. The sender generates the chaptern[] or chapterc[] bitfields, by looping through the state arrays. If the note_seqnum[] or control_seqnum[] value for an array index is non-zero, the sender examines the associated note_state[] or control_state[] array element, and codes data from the element into the bitfield. After the looping completes, the sender sets the chapter size variable to code the final bitfield length. In addition, the sender clears the S bit of the chapterp[], chapterw[], or chapterc[] bitfield. For chaptern[], the sender clears the S bit or the B bit of the bitfield, as described in Appendix A.4 of [1]. Next, the sender refreshes the upper levels of the RJSS hierarchy. At the second-level, the sender updates the cheader[] bitfield of the channel associated with the command. The sender sets the S bit of cheader[] to 0. If the new command forced the addition of a new chapter or channel journal, the sender may also update other cheader[] fields. At the top-level, the sender updates the top-level jheader[] bitfield in a similar manner. Finally, the sender updates the seqnum variables associated with the changed bitfield arrays. The sender sets the seqnum variables to the extended sequence number of the packet. Lazzaro/Wawrzynek [Page 22] INTERNET-DRAFT 15 January 2003 5.4 Trimming the RJSS At regular intervals, receivers send RTCP receiver reports to the sender (as described in Section 6.4.2 of [2]). These reports include the extended highest sequence number received (EHSNR) field. This field codes the highest sequence number that the receiver has observed from the sender, extended to disambiguate sequence number rollover. When the sender receives an RTCP receiver report, it runs the RJSS trimming algorithm. The trimming algorithm uses the EHSNR to trim away parts of the RJSS, and thus reduce the size of recovery journals sent in subsequent RTP packets. The algorithm (as applied to a two-party session) relies on the following observation: if the EHSNR indicates that a packet with sequence number K has been received, MIDI commands sent in packets with sequence numbers I <= K may be removed from the RJSS without violating the recovery journal mandate defined in Section 4 of [1]. To begin the trimming algorithm, the sender extracts the EHSNR field from the receiver report, and adjusts the EHSNR to reflect the sequence number extension prefix of the sender. Then, the sender compares the adjusted EHSNR value with seqnum fields at each level of the RJSS, starting at the top level. Levels whose seqnum is less than or equal to the adjusted EHSNR are trimmed, by setting the seqnum to zero. If necessary, the jheader[] and cheader[] arrays above the trimmed level are adjusted to match the new journal layout. The checkpoint packet sequence number field of jheader[] is updated to match the EHSNR. At the leaf level, the sender trims the size of the variable-length chaptern[] and chapterc[] bitfields. The sender loops through the note_seqnum[] or control_seqnum[] array, and clears elements whose value is less than or equal to the adjusted EHSNR. The sender then creates a new chaptern[] or chapterc[] bitfield, and updates the LENGTH field of the associated cheader[] bitfield. Note that the trimming algorithm does not add information to the checkpoint history. As a consequence, the trimming algorithm does not clear the S bit (and for chaptern[], the B bit) of any recovery journal bitfield. As a second consequence, the trimming algorithm does not set RJSS seqnum variables to the EHSNR value. Lazzaro/Wawrzynek [Page 23] INTERNET-DRAFT 15 January 2003 5.5 Implementation Notes For clarity, the recovery journal sender implementation we describe has been simplified in several ways. In this section, we discuss the improvements that would be found in a complete, efficient sender implementation suitable for use in a production system. In a production implementation, the sending structure shown in Figure 9 would be modified to cover the full recovery journal syntax. Chapter journal structures would be added for the missing channel and system chapters defined in Appendices A and B of [1]. An efficient implementation would use enhanced versions of the traversing, updating, and trimming algorithms presented in Sections 5.2-4. In particular, the Chapter N and Chapter C algorithms would use more sophisticated RJSS data structures, in order to avoid looping through all 128 note or controller numbers. The recovery journal sender implemented in [19] includes enhancements of this type. Finally, a production sender implementation would probably implement algorithms that support a variety of MWPP application domains (two-party topologies and multi-party topologies, RTCP and no-RTCP, etc). In the Appendices of this memo, we discuss recovery journal sender issues for application domains beyond the two-party example system described above. Lazzaro/Wawrzynek [Page 24] INTERNET-DRAFT 15 January 2003 6. Receiving MWPP Streams: General Considerations In this section, we discuss MWPP receiver implementation issues, in the context of the interactive session introduced in Section 2. To begin, we imagine that an ideal network carries the RTP stream. Packets are never lost or reordered, and the end-to-end latency is constant. In addition, we assume that all MIDI commands coded in the MIDI list of a packet share the same command execution timestamp (as defined in Section 3 of [1]), and that the default semantics for command timestamps are in effect. Under these conditions, a simple algorithm may be used to render a high- quality performance. Upon the receipt of an RTP packet, the receiver immediately executes the commands coded in the MIDI command section of the payload. Commands are executed in order of their appearance in the MIDI list. The command timestamps are ignored. Unfortunately, this simple algorithm breaks down once we relax our assumptions about the network and the MIDI list: 1. If we permit lost and reordered packets to occur in the network, the algorithm may produce unrecoverable rendering artifacts, violating the mandate defined in Section 4 of [1]. 2. If we permit the network to exhibit variable latency, the algorithm modulates the network jitter onto rendered MIDI command stream. 3. If we permit a MIDI list to code commands with different timestamps, the algorithm adds temporal jitter to the rendered performance, as it ignores MIDI list timestamps. In this section, we discuss interactive receiver design techniques under these relaxed assumptions (see Appendix A for a discussion of content streaming receiver design). Interactive receiver design is not a "one size fits all" endeavor. Applications often target specific types of network environments, and receiver algorithms are crafted to work well on those networks. In the sections below, we describe a complete receiver design for high- performance WAN networks (Section 6.1) and discuss design issues for other types of networks (Section 6.2). Lazzaro/Wawrzynek [Page 25] INTERNET-DRAFT 15 January 2003 6.1 The Network Musical Performance (NMP) Receiver Design In this section, we describe the MWPP receiver implemented in the Network Music Performance (NMP) system described in [13] and implemented in [19]. The NMP system is an interactive musical performance application that uses an early prototype version of MWPP. Musicians located at different sites interact over the network to perform as they would if located in the same room, using MIDI controllers as instruments. NMP is designed for use between university sites within the State of California in the USA, using the CalREN2 network. In an NMP session, network artifacts may affect how a musician hears the MIDI performances of remote players. However, the network does not affect how a musician hears his own performance. In this way, NMP differs from MWPP LAN applications. In LAN work, a musician usually hears his own MIDI performance via the network link. Several aspects of CalREN2 network behavior (as measured in 2001 timeframe, as documented in [13]) guided the NMP system design: o The median symmetric latency (1/2 the round-trip time) of packets sent between network sites is comparable to the acoustic latency between two musicians located in the same room. For example, the latency between Berkeley and Stanford is 2.1 ms, corresponding to an acoustic distance of 2.4 feet (0.72 meters). These campuses are 40 miles (64 km) apart. o For most times of day, the nominal temporal jitter is quite short (for Berkeley-Stanford, the standard deviation of the round-trip time was under 200 microseconds). o For most times of day, a few percent (0-4%) of the packets sent arrive significantly late (> 40 ms), probably due to a queuing transient somewhere in the network path. More rarely (< 0.1%), a packet is lost during the transient. o At predictable times during the day (before lunchtime, at the end of the workday, etc), network performance deteriorates (10-20% late packets) in a manner that makes the network unsuitable for low-latency interactive use. o CalREN2 has deeply over-provisioned bandwidth, relative to MIDI bandwidth usage. Lazzaro/Wawrzynek [Page 26] INTERNET-DRAFT 15 January 2003 The NMP sender freely uses network bandwidth to improve the performance experience. As soon as a musician generates a MIDI command, an RTP packet coding the command is sent to the other players. This sending algorithm reduces latency at the cost of bandwidth. In addition, guard packets (described in Section 4.2) are sent at frequent intervals, to minimize the impact of packet loss. The NMP receiver maintains a model of the stream, and uses this model as the basis of its resiliency system. Upon the receipt of an MWPP packet, the receiver predicts the RTP sequence number and the RTP timestamp (with error bars) of the packet. Under normal network conditions, about 95% of received packets fit the predictions [13]. In this common case, the receiver immediately executes the MIDI command coded in the packet. Note that the NMP receiver does not use a playout buffer; the design is optimized for lowest latency at the expense of command jitter. Occasionally, an incoming packet fits the sequence number prediction but falls outside the timestamp prediction error bars (see Appendix B of [13] for timestamp model details). In most cases, the receiver still executes the MIDI command coded in the packet. An important exception is MIDI NoteOn commands with non-zero velocity: the receiver discards these commands. By discarding late commands that sound notes, the receiver prevents "straggler notes" from disturbing a performance. By executing all other late MIDI commands, the receiver quiets "soft stuck notes" immediately, and updates all other MIDI state in an acceptable way. More rarely, an incoming packet does not fit the sequence number prediction. The receiver keeps track of the highest sequence number received in the stream, and predicts that an incoming packet will have a sequence number one greater than this value. If the sequence number of an incoming packet is greater than the prediction, a packet loss has occurred. If the sequence number of the received packet is less than the prediction, the packet has been received out of order. All sequence number calculations are modulo 2^16, and use standard methods (described in [2]) to avoid tracking errors during rollover. If a packet loss has occurred, the receiver examines the journal section of the received packet, and uses it to gracefully recover from the loss episode. We describe this recovery procedure in Section 7 of this memo. The recovery process may result in the execution of one or more MIDI commands. After executing the recovery commands, the receiver processes the MIDI command encoded in the packet, using the timestamp model test described above. Lazzaro/Wawrzynek [Page 27] INTERNET-DRAFT 15 January 2003 If a packet is received out of order, the receiver ignores the packet. The receiver takes this action because a packet received out of order is always preceded by a packet that signalled a loss event. This loss event triggered the recovery process, which may have executed recovery commands. The MIDI command coded in the out-of-order packet might, if executed, duplicate these recovery commands, and this duplication might endanger the integrity of the stream. Thus, ignoring the out-of-order packet is the safe approach. 6.2 Receiver Design Issues The NMP receiver targets a network with a particular set of characteristics: low nominal jitter, low packet loss, and occasional outlier packets that arrive very late. In this section, we consider how networks with different characteristics impact MWPP receiver design. Networks with significant nominal jitter cannot use the buffer-free receiver design described in Section 6.1. For example, the NMP system performs poorly for musicians that use dial-up modem connections, because the buffer-free receiver design modulates modem jitter onto the performances. Receivers designed for high-jitter networks should use a playout buffer. References [17] and [20] describe how to use playout buffers in latency-critical applications. Appendix A.2 may also be interest, as it addresses MWPP-specific playout buffer issues. Receivers intended for use on LANs face a different set of issues. A dedicated LAN fabric built with modern hardware is in many ways a predictable environment. The network problems addressed by the NMP receiver design (packet loss and outlier late packets) might only occur under extreme network overload conditions. Systems designed for this environment may choose to configure streams without the recovery journal system (Appendix C.1.1 of [1]). Receivers may also wish to forego, or simplify, the detection of outlier late packets. Receivers should monitor the RTP sequence numbers of incoming packets, to detect network unreliability. However, in some respects, LAN applications may be more demanding than WAN applications. In LAN applications, musicians may be receiving performance feedback from audio that is rendered from the MWPP stream. The tolerance a musician has for latency and jitter in this context may be quite low. To reduce the perceived jitter, receivers may use a small playout buffer (in the range of 100us to 2ms). The buffer does add a a small amount of latency to the system, that may be annoying to some players. Receiver designs should include buffer tuning parameters, to let musicians adjust the tradeoff between latency and jitter. Lazzaro/Wawrzynek [Page 28] INTERNET-DRAFT 15 January 2003 7. Receiving MWPP Streams: The Recovery Journal In this section, we describe the recovery algorithm used by the NMP receiver [13]. In most ways, the recovery techniques we describe are generally applicable to interactive MWPP receiver design. However, a few aspects of the design are specialized for the NMP system: o The recovery algorithm covers the subset of MIDI commands used by MPEG 4 Structured Audio [7]. Structured Audio does not use use MIDI Systems (0xF) commands, and uses MIDI Control Change (0xB) commands in a simplified way. o The NMP system does not use a playout buffer, and so the recovery algorithm does not address interactions with a playout buffer. In addition, to simplify the discussion, we omit receiver support for the Poly Aftertouch (0xA) and Channel Aftertouch (0xD) voice commands. At a high level, the receiver algorithm works as follows. Upon the detection of a packet loss, the receiver examines the recovery journal of the packet that ends the loss event. If necessary, the receiver executes one or more MIDI commands to recover from the loss. To prepare for recovery, a receiver maintains a data structure, the Recovery Journal Receiver Structure (RJRS). The RJRS codes information about the MIDI commands the receiver executes (both incoming stream commands and self-generated recovery commands). At the start of the stream, the RJRS is initialized to code that no commands have been executed. Immediately after executing a MIDI command, the receiver updates the RJRS with information about the command. We now describe the recovery algorithm in detail. We begin with two definitions that classify loss events. These definitions assume that the packet that ends the loss event has RTP sequence number I. o Single-packet loss. A single-packet loss occurs if the last packet received before the loss event (excluding out-of-order packets) has the sequence number I-2 (modulo 2^16). o Multi-packet loss. A multi-packet loss occurs if the last packet received before the loss event (excluding out-of-order packets) has a sequence number less than I-2 (modulo 2^16). Lazzaro/Wawrzynek [Page 29] INTERNET-DRAFT 15 January 2003 Upon the detection of a packet loss, the recovery algorithm begins by examining the recovery journal header (Figure 7 of [1]), to check for several special cases: o If the header field A is 0, the recovery journal has no channel journals, and so no action is taken. Note that if this algorithm supported MIDI Systems commands, it would also examine the Y field. o If a single-packet loss has occurred, and the header S bit is 1, the lost packet has a MIDI command section with an empty MIDI list. No action is taken. If these checks fail, the recovery algorithm proceeds to parse the recovery journal body. For each channel journal (Figure 8 in [1]) in the recovery journal, the receiver compares the data in each chapter journal (Appendix A of [1]) to the RJRS data for the chapter. If the data are inconsistent, the algorithm infers that MIDI command(s) related to the chapter journal have been lost. The recovery algorithm executes MIDI commands to repair this loss, and updates the RJRS to reflect the repair. For multi-packet losses, the receiver parses each channel and chapter journal and checks for inconsistency. For single-packet losses, journal parsing is more efficient, as the receiver may skip channel and chapter journals whose S bits are set to 1. If the NMP recovery algorithm had supported MIDI System commands, the system chapters (Appendix B in [1]) of the system journal (Figure 9 in [1]) would be compared to systems data stored in the RJRS. If the recovery algorithm discovered inconsistency, MIDI System commands would be executed to repair the loss. In the sections that follow, we describe the recovery steps that are specific to each chapter journal. We also describe how to update the RJRS for the command types associates with the chapter journal. We cover 4 chapter journal types: W (Pitch Wheel, 0xE), N (Note, 0x8 and 0x9), C (Control Change, 0xB) and P (Program Change, 0xC). Chapters are parsed in the order of appearance in the Table of Contents of the channel journal header (P, then W, then N, then C). The sections below reference the C implementation of the RJRS shown in Figure 10. This structure is hierarchical, reflecting the recovery journal architecture. At the leaf level, specialized data structures (jrec_chapterw, jrec_chaptern, jrec_chapterc, and jrec_chapterp) code state variables for a single chapter journal type. A mid-level structure (jrec_channel) represents a single MIDI channel, and a top-level structure (jrec_stream) represents the entire MIDI stream. Lazzaro/Wawrzynek [Page 30] INTERNET-DRAFT 15 January 2003 typedef unsigned char uint8; /* must be 1 octet */ typedef unsigned short uint16; /* must be 2 octets */ typedef unsigned long uint32; /* must be 4 octets */ /***********************************************************/ /* leaf level of hierarchy: Chapter W, Appendix A.3 of [1] */ /***********************************************************/ typedef struct jrec_chapterw { /* MIDI Pitch Wheel (0xE) */ uint16 val; /* most recent 14-bit wheel value */ } jrec_chapterw; /***********************************************************/ /* leaf level of hierarchy: Chapter N, Appendix A.4 of [1] */ /***********************************************************/ typedef struct jrec_chaptern { /* Note commands (0x8, 0x9) */ /* arrays of length 128 --> one for each MIDI Note number */ uint32 time[128]; /* exec time of most recent NoteOn */ uint32 extseq[128]; /* extended seqnum for that NoteOn */ uint8 vel[128]; /* NoteOn velocity (0 for NoteOff) */ } jrec_chaptern; /***********************************************************/ /* leaf level of hierarchy: Chapter C, Appendix A.7 of [1] */ /***********************************************************/ typedef struct jrec_chapterc { /* Control Change (0xB) */ /* array of length 128 --> one for each controller number */ uint8 value[128]; /* Chapter C value tool state */ uint8 count[128]; /* Chapter C count tool state */ uint8 toggle[128]; /* Chapter C toggle tool state */ } jrec_chapterc; Figure 10 -- Recovery Journal Receiving Structure (part 1) Lazzaro/Wawrzynek [Page 31] INTERNET-DRAFT 15 January 2003 /***********************************************************/ /* leaf level of hierarchy: Chapter P, Appendix A.2 of [1] */ /***********************************************************/ typedef struct jrec_chapterp { /* MIDI Program Change (0xC) */ uint8 prognum; /* most recent 7-bit program value */ uint8 prognum_qual; /* 1 once first 0xC command arrives */ uint8 coarse; /* most recent bank coarse value */ uint8 coarse_qual; /* 1 once first 0xBn 0x00 arrives */ uint8 fine; /* most recent bank fine value */ uint8 fine_qual; /* 1 once first 0xBn 0x20 arrives */ } jrec_chapterp; /***************************************************/ /* second-level of hierarchy, for MIDI channels */ /***************************************************/ typedef struct jrec_channel { jrec_chapterw chapterw; /* Pitch Wheel (0xE) info */ jrec_chaptern chaptern; /* Note (0x8, 0x9) info */ jrec_chapterp chapterp; /* Program Change (0xC) info */ jrec_chapterc chapterc; /* Control Change (0xB) info */ } jrec_channel; /***********************************************/ /* top level of hierarchy, for the MIDI stream */ /***********************************************/ typedef struct jrec_stream { jrec_channel channels[16]; /* index is MIDI channel */ } jrec_stream; Figure 10 (continued) -- Recovery Journal Receiving Structure Lazzaro/Wawrzynek [Page 32] INTERNET-DRAFT 15 January 2003 7.1 Chapter W: MIDI Pitch Wheel (0xE) Chapter W of the recovery journal protects against the loss of MIDI Pitch Wheel (0xE) commands. A common use of the Pitch Wheel command is to transmit the current position of a "pitch wheel" controller placed on the side of MIDI piano controllers. Players use the pitch wheel to dynamically alter the pitch of all depressed keys. The NMP receiver maintains the jrec_chapterw structure (Figure 10) for each voice channel in jrec_stream, to code pitch wheel state information. In jrec_chapterw, val holds the 14-bit data value of the most recent Pitch Wheel command that has arrived on a channel. At the start of the stream, val is initialized to the default pitch wheel value (0x2000). The NMP receiver uses jrec_chapterw in its recovery algorithm. While parsing the recovery journal, it may find a Chapter W (Appendix A.3 in [1]) bitfield in a channel journal. This chapter codes the 14-bit data value of the most recent MIDI Pitch Wheel command in the checkpoint history. If the Chapter W and jrec_chapterw pitch wheel values do not match, one or more commands have been lost. To recover from this loss, the NMP receiver immediately executes a MIDI Pitch Wheel command on the channel, using the data value coded in the recovery journal. The receiver then updates the jrec_chapterw variables to reflect the executed command. 7.2 Chapter N: MIDI NoteOn (0x8) and NoteOff (0x9) Chapter N of the recovery journal protects against the loss of MIDI NoteOn (0x9) and NoteOff (0x8) commands. In this section, we consider NoteOn commands with a velocity value of 0 to be NoteOff commands. If an unprotected NoteOn command is lost, a note is skipped. If an unprotected NoteOff command is lost, a note may sound indefinitely. The NMP receiver maintains the jrec_chaptern structure (Figure 10) for each voice channel in jrec_stream, to code note-related state information. State is kept for each of the 128 note numbers on a channel, using three arrays of length 128 (vel[], seq[], and time[]). The elements of these arrays are initialized to zero at the start of a stream. The vel[n] array element holds information about the most recent note command for note number n. If this command is a NoteOn command, vel[n] holds the velocity data for the command. If this command is a NoteOff command, vel[n] is set to 0. The time[n] and extseq[n] array elements code information about the most recently executed NoteOn command. Lazzaro/Wawrzynek [Page 33] INTERNET-DRAFT 15 January 2003 The time[n] element holds the execution time of the command, referenced to the local timebase of the receiver. The extseq[n] element holds the RTP extended sequence number of the packet associated with the command. For incoming stream commands, extseq[n] codes the packet of the associated MIDI list. For recovery commands, extseq[n] codes the packet of the associated recovery journal. The NMP receiver uses the jrec_chaptern state information in its recovery algorithm. The Chapter N recovery journal bitfield (Figure A.4.1 in [1]) consists of two data structures: a bit array coding recently-sent NoteOff commands that are vulnerable to packet loss, and a note log list coding recently-sent NoteOn commands that are vulnerable to packet loss. Recovery processing begins with the NoteOff bit array. For each set bit in the array, the receiver checks the corresponding vel[n] element in jrec_chaptern. If vel[n] is non-zero, a NoteOff command, or a NoteOff->NoteOn->NoteOff command sequence, has been lost. To recover from this loss, the receiver immediately executes a NoteOff command for the note number on the channel, and sets vel[n] to 0. The receiver then parses the note log list. For each NoteOn log in the list, the receiver checks the corresponding vel[n] element. If vel[n] is zero, a NoteOn command, or a NoteOn->NoteOff->NoteOn command sequence, has been lost. The receiver may execute the most recent lost NoteOn (to play the note) or may take no action (to skip the note), based on criteria we describe at the end of this section. Whether the note is played or skipped, the receiver updates the vel[n], time[n], and extseq[n] elements as if the NoteOn executed. If vel[n] is non-zero, the receiver performs several checks to test if a NoteOff->NoteOn sequence has been lost. o If vel[n] does not match the note log velocity, the note log must code a different NoteOn command, and thus a NoteOff->NoteOn sequence has been lost. o If extseq[n] is less than the (extended) checkpoint packet sequence numbed coded in the recovery journal header (Figure 7 of [1]), the vel[n] NoteOn command is not in the checkpoint history, and thus a NoteOff->NoteOn sequence has been lost. o If the Y bit is set to 1, the NoteOn is musically "simultaneous" with the RTP timestamp of the packet. If time[n] codes a time value that is clearly not recent, a NoteOff->NoteOn sequence has been lost. Lazzaro/Wawrzynek [Page 34] INTERNET-DRAFT 15 January 2003 If these tests indicate a lost NoteOff->NoteOn sequence, the receiver immediately executes a NoteOff command. The receiver decides if the most graceful action is to play or to skip the lost NoteOn, using the criteria we describe at the end of this section. Whether or not the receiver issues a NoteOn command, the vel[n], time[n], and extseq[n] arrays are updated as if it did. Note that the tests above do not catch all lost NoteOff->NoteOn commands. If a fast NoteOn->NoteOff->NoteOn sequence occurs on a note number, with identical velocity values for both NoteOn commands, a lost NoteOff->NoteOn does not result in the recovery algorithm generating a NoteOff command. Instead, the first NoteOn continues to sound, to be terminated by the future NoteOff command. In practice, this (rare) outcome is not musically objectionable. Finally, we discuss how the receiver decides whether to play or to skip a lost NoteOn command. The note log Y bit is set if the NoteOn is "simultaneous" with the RTP timestamp of the packet holding the note log. If Y is 0, the receiver does not execute a NoteOn command. If Y is 1, and if the packet has not arrived late, the receiver immediately executes a NoteOn command for the note number, using the velocity coded in the note log. 7.3 Chapter C: MIDI Control Change (0xB) Chapter C (Appendix A.7 in [1]) protects against the loss of MIDI Control Change commands. A Control Change command alters the 7-bit value of one of the 128 MIDI controllers. Chapter C offers three tools for protecting a Control Change command: the value tool (for graded controllers such as sliders) the toggle tool (for on/off switches) and the count tool (for momentary-contact switches). Senders choose a tool to encode recovery information for a controller, and encode the tool type along with the data in the journal (Figures A.7.2 and A.7.3 in [1]). A few uses of Control Change commands are not solely protected by Chapter C. The protection of controllers 0 and 32 (Bank Coarse and Bank Fine) is shared between Chapter C and Chapter P (Section 7.4). In addition, some controllers are used to implement a system for setting secondary parameters (the Registered Parameter Number (RPN) and the Non- Registered Parameter Number (NRPN) systems). Chapter M (Appendix A.8 of [1]) protects the RPN and NRPN system. MPEG 4 Structured Audio [7] does not use these systems, and so the NMP system does not use Chapter M. Lazzaro/Wawrzynek [Page 35] INTERNET-DRAFT 15 January 2003 The NMP receiver maintains the jrec_chapterc structure (Figure 10) for each voice channel in jrec_stream, to code Control Change state information. The value[] array holds the most recent data values for each controller number. At the start of the stream, value[] is initialized to the SA default controller data values specified in [7]. The count[] and toggle[] arrays hold the count tool and toggle tool state values. At the start of a stream, these arrays are initialized to zero. Whenever a Control Command executes, the receiver updates the count[] and toggle[] state values, using the algorithms described in Appendix A.7 of [1]. The NMP receiver uses the jrec_chapterc state information in its recovery algorithm. The Chapter C bitfield consists of a list of controller logs. Each log codes the controller number, the tool type, and the state value for the tool. For the log for controller number n, the receiver determines the tool type in use (value, toggle, or count), and compares the data in the log to the associated jrec_chapterc array element (value[n], toggle[n], or count[n]). If the data do not match, one or more Control Change commands have been lost. The method the NMP receiver uses to recover from this loss depends on the tool type and the controller number. For graded controllers protected by the value tool, the receiver executes a Control Change command using the new data value. For the toggle and count tools, the recovery action is more complex. For example, the Hold Pedal (64) controller is typically used as a sustain pedal for piano-like sounds, and is typically coded using the toggle tool. If Hold Pedal Control Change command(s) are lost, the NMP receiver takes different actions depending on the starting and ending state of the lost sequence, to ensure "ringing" piano notes are "damped" to silence. After recovering from the loss, the receiver updates the value[], toggle[], and count[] arrays to reflect the Chapter C data and the executed commands. 7.4 Chapter P: MIDI Program Change (0xC) Chapter P of the recovery journal protects against the loss of MIDI Program Change (0xC) commands. A common use for Program Change commands is to select the timbre of a channel. The 7-bit data value of the command selects one of 128 possible timbres. The binding of data values to instrument timbres is managed by the rendering algorithm in use. Lazzaro/Wawrzynek [Page 36] INTERNET-DRAFT 15 January 2003 To increase the number of possible timbres, MIDI Control Change (0xB) commands may be issued prior to the Program Change command, to select which "bank" of programs is in use. The Bank Coarse (controller number 0) and Bank Fine (controller number 32) Control Change commands may be used together, to specify the 14-bit bank number that subsequent Program Change commands reference. Alternatively, the Bank Coarse controller number may be used alone to specify a 7-bit bank number. The NMP receiver maintains the jrec_chapterp structure (Figure 10) for each voice channel in jrec_stream, to code Program Change state information. The prognum variable of jrec_chapterp holds the data value for the most recent Program Change command that has arrived on the stream. The coarse and fine variables of jrec_chapterp code the Bank Coarse and Bank Fine Control Change data values that were in effect when that Program Change command arrived. The prognum_qual, coarse_qual and fine_qual variables are initialized to 0, and are set to 1 upon the receipt of the first Program Change, Bank Coarse Control Change, and Bank Fine Control Change command, respectively. The NMP receiver uses jrec_chapterp in its recovery algorithm. While parsing the recovery journal, it may find a Chapter P (Appendix A.2 in [1]) bitfield in a channel journal. Fields in Chapter P code the data value for the most recent Program Change command, and the coarse and fine bank values in effect for that Program Change command (if any). The receiver checks to see if these recovery journal fields match the data stored in jrec_chapterp. If these checks fail, one or more Program Change commands have been lost. To recover from this loss, the receiver takes the following steps. If the C (coarse) or F (fine) bits in Chapter P are set (Figure A.2.1 in [1]), Control Change bank command(s) have preceded the Program Change command. The receiver compares the bank data coded by Chapter P with the current bank data for the channel (coded in jrec_channelc). If the bank data do not agree, the receiver issues Control Change command(s) to align the stream with Chapter P. The receiver then updates jrec_channelp and jrec_channelc variables to reflect the executed command(s). Finally, the receiver issues a Program Change command that reflects the data in Chapter P, and updates the prognum and qual_prognum fields in jrec_channelp. Note that this method relies on Chapter P recovery to precede Chapter C recovery during channel journal processing. This ordering ensures that lost bank select Control Change that occur after a lost Program Change command in a stream are handled correctly during Chapter C parsing. Lazzaro/Wawrzynek [Page 37] INTERNET-DRAFT 15 January 2003 8. Congestion Control Congestion control issues for MWPP implementations are discussed in detail in Section 4.3 of this memo. Also see Section 8 of [1]. 9. Security Considerations General security considerations for MWPP are discussed in detail in Section 7 of [1]. Supplemental discussion on MWPP implementation security issues is presented in Section 3 of this memo. 10. Acknowledgments See the Acknowledgments section of [1]. Lazzaro/Wawrzynek [Page 38] INTERNET-DRAFT 15 January 2003 Appendix A. Content Streaming with MWPP In this Appendix, we show how to use a media server to distribute MIDI performances to one or more clients. We refer to applications of this type as content streaming applications. The content source may be a live MIDI concert, a pre-recorded MIDI file, or a dynamically-generated MIDI stream that slowly changes in response to client user activity (such as clicks on web links). Interactive and content-streaming MWPP applications differ in the role of latency in the application. Interactive applications place the network in the sensory-motor loop of a single musician, or in the performance loop between several musicians. To optimize the user experience, these applications run at or near the underlying latency of the network. Receivers use a minimal playout buffer (or no playout buffer at all), and rely on the specialized methods for lost and late packet recovery described in the main text. In comparison, clients and servers in content-streaming applications interact at a relatively slow time constant. As a consequence, MWPP clients may use a playout buffer to smooth network jitter, without impacting the user response time. Clients may also use the playout buffer in conjunction with generic forward-error correction (FEC, [12]) or packet retransmission [21] in order to replace lost packets. In the sections below, we describe how to set up MWPP content streaming sessions (Appendix A.1), discuss baseline client and server streaming algorithms (Appendix A.2), and show how to enhance MWPP content streaming with an ancillary packet replacement stream (Appendix A.3). A.1 Content Streaming: Session Management In this Appendix, we show how content-streaming servers and clients set up MWPP streams. We assume the participants use the Real Time Streaming Protocol (RTSP, [6]) to manage the session. Like HTTP, RTSP identifies a media stream with a URL. For example, the RTSP URL rtsp://cs.example.net/ode_to_joy may identify a MWPP stream of a Bach performance. By default, an RTSP URL implies that an RTSP server may be accepting TCP connections at port 554 of the host name that follows the double slash in the RTSP URL. However, note that RTSP may also use UDP (rtspu://) or TLS (rtsps://) transport. In a typical use, a client initiates contact with the server. For a TCP RTSP URL, the client opens a TCP connection to the RTSP server named in the URL, and sends the server a series of RTSP messages to set up the session. Lazzaro/Wawrzynek [Page 39] INTERNET-DRAFT 15 January 2003 v=0 o=server 2520644554 2838152170 IN IP4 server.example.net s=Example t=0 0 a=recvonly m=audio 0 RTP/AVP 61 a=control:rtsp://cs.example.net/ode_to_joy/baseline a=rtpmap: 61 mpeg4-generic/44100 a=fmtp: 61 streamtype=5; mode=mwpp; config=""; profile-level-id=76; a=fmtp: 61 render=sasc; inline="e4"; compr=none; Figure A.1 -- RTSP session description One RTSP method, DESCRIBE, returns the SDP session description associated with the RTSP URL. Figure A.1 shows a session description returned by RTSP for an example MWPP session. The session description presents the session from the view of the client (note the recvonly attribute). Like the session descriptions for interactive applications shown in the main text, the RTSP session description in Figure A.1 codes media initialization information for the MWPP session. In this example, the session uses an mpeg4-generic MWPP stream to specify a General MIDI renderer. However, in most cases, RTSP session descriptions do not code explicit transport information in the session description (transport type, network addresses, port numbers, etc). Instead, RTSP session descriptions usually code RTSP URLs that may be used to negotiate transport details for each stream. The URLs appear as control attributes in the session description (for example, in Figure A.1 rtsp://cs.example.net/ode_to_joy/baseline). Clients use the RTSP SETUP method to translate an RTSP control URL into concrete transport information. This indirect approach supports flexible transport setup, and is useful for working around network middleboxes (such as NATs and firewalls). To overcome stubborn network obstacles, RTSP supports interleaving RTP and RTCP streams over the TCP connection that carries RTSP message traffic. Clients also use RTSP methods to start (the PLAY method) and stop (the PAUSE method) media flow. The PLAY method supports parameters that specify the starting point of the stream, and thus together with PAUSE implements the full set of tape-deck remote control commands (rewind, fast-forward, play, and pause). To end a session, the client uses the TEARDOWN method. Lazzaro/Wawrzynek [Page 40] INTERNET-DRAFT 15 January 2003 Once the RTSP server replies to the SETUP method, the client sets up the RTP and RTCP stream(s) for the session. For UDP media streams, the client may use the interactive setup algorithms described in Sections 2 and 3 of the main text. For TCP media streams that establish separate connections for media flow, the client may follow the interactive MWPP TCP guidance in Appendix C.1 and C.2. For media streams that interleave RTP and RTCP into the RTSP TCP connection, the client may follow the guidance in Appendix C.3. A.2 Content Streaming: Baseline Algorithms In this Appendix, we describe MWPP sending and receiving algorithms for content-streaming sessions. We focus on MWPP streaming over unicast UDP transport. A client in an MWPP content-streaming application implements an MWPP receiver. Unlike the interactive receiver described in Sections 6 and 7 of the main text, a content-streaming receiver implements a playout buffer. Below, we present a brief sketch of a simple receiver design, to introduce the first-order design issues. The heart of the receiver design is the playout buffer. As the heart pumps blood, the playout buffer pumps packets. Architecturally, an MWPP playout buffer is a queue of pointers to MWPP packets, ordered by sequence number (lowest numbers at the front of the queue, modulo 2^32). At the start of a stream, the queue is empty. As packets arrive on the RTP port, the receiver places them at the back of the queue. If the sequence number of a new packet indicates a loss event, the receiver adds empty slots to the back of the queue for the lost packets, and places the new packet behind the empty slots. If a packet arrives out of order, the receiver places the packet into the empty queue slot reserved for it. The receiver renders audio by removing MWPP packets from the front of the queue. At the start of a stream, the receiver does not start rendering the queue immediately. Instead, the receiver waits until the stream time (as determined by the RTP timestamps) held in the queue matches the desired buffer latency. Receivers choose the buffer latency to match the requirements of the application, balancing user interaction response time (aided by low buffer latency) and the fidelity of the rendered stream (aided by high buffer latency). We now describe how the receiver removes packets from the front of the queue. To simplify the explanation, we assume that the MIDI command section of each packet holds one MIDI command, whose execution time is coded by RTP timestamp of the packet. In practice, well-encoded packets Lazzaro/Wawrzynek [Page 41] INTERNET-DRAFT 15 January 2003 in content-streaming applications will hold several (or perhaps many) MIDI commands. However, the extension of the rendering algorithm we present to multi-command packets is straightforward. To start rendering the queue, a receiver takes the first packet off the queue, and initializes a variable time_pointer to the RTP timestamp value of the packet. The receiver extracts the MIDI command from the packet, and passes the command to the MIDI rendering system for immediate execution. At regular intervals thereafter (say, once every 250 microseconds), the receiver increments time_pointer by the interval value, and checks if the packet closest to the front of the queue has an RTP timestamp that is less than or equal to time_pointer. If so, and if the packet is at the very front of the queue, the receiver takes the packet off the queue, and extracts the MIDI command from the packet. However, if an empty slot is at the front of the queue, a packet loss event has occurred. The receiver removes the empty queue slot(s), removes the packet that follows the queue slots, and uses the recovery journal techniques described in Section 7 of the main text to restore stream integrity. Then, the receiver extracts the MIDI command from the MIDI command section of the packet, Finally, the receiver checks to see if the command timestamp of the extracted MIDI command is reasonably close in time to time_pointer. If so, the receiver passes the command to the rendering system for immediate execution. If the timestamp check reveals a late command, the receiver uses the heuristics described in Section 6.1 of the main text to decide whether to execute or skip the command. We now turn our attention to server design issues. A server in an MWPP content-streaming application implements an MWPP sender. However, a content-streaming MWPP sender differs in several ways from the interactive sender described in Sections 4 and 5 in the main text. Unlike interactive receivers, content-streaming receivers use a playout buffer, and rely on RTP timestamps to schedule MIDI command execution. Receivers with these characteristics work best if senders (1) generate MWPP packets that code a constant interval of media time and (2) transfer these packets at a relatively stable rate. Senders may use the SDP ptime [5] parameter to indicate the approximate duration of the MWPP packets it sends. Note that the exact duration may vary from packet to packet, due to the event-based nature of MIDI. Lazzaro/Wawrzynek [Page 42] INTERNET-DRAFT 15 January 2003 In choosing the packet duration, senders balance several issues. A longer duration improves efficiency, as header overhead is amortized over longer time periods. However, a shorter duration reduces the perceptual impact of single packet loss. In addition, a shorter duration extends the range of possible playout buffer latencies to smaller values. A.3 Content Streaming: Packet Replacement Streams In this Appendix, we describe how applications may replace lost packets in an MWPP stream, by using redundant data sent on an ancillary stream in the session. Applications use this technique to improve the fidelity of a rendered performance, by avoiding the use of the recovery journal system for minor loss events. As described in Appendix A.2, the playout buffer of a receiver is organized as a queue of pointers to incoming RTP packets. If a loss event occurs, the receiver leaves empty slots in the queue for the lost packets. In this Appendix, we describe how receivers may use an ancillary stream to fill the empty queue slots with replacement packets. For packet replacement to be effective, empty slots must be filled before they reach the front of the queue (and playout occurs). Our examples use payload-independent tools [12] [21] for RTP packet replacement, as MWPP does not define MIDI-specific redundancy tools. One RTP tool for MWPP packet replacement is generic forward error correction (FEC, [12]). [12] describes a feed-forward system that does not use receiver feedback. In this system, an ancillary stream carries an encoded redundant copy of the primary MWPP stream. If a packet loss occurs on the primary stream, the receiver attempts to reconstruct the lost packet by processing the ancillary stream. Figure A.2 shows an MWPP session description that adds an FEC stream to the MWPP primary stream. The RTP stream with PTYPE number 62 carries the FEC stream, using the ulpfec format defined in [12]. If a client wishes to receive the FEC stream, it uses the RTSP URL rtsp://cs.example.net/ode_to_joy/fec to set up the stream. Lazzaro/Wawrzynek [Page 43] INTERNET-DRAFT 15 January 2003 v=0 o=server 2520644554 2838152170 IN IP4 server.example.net s=Example t=0 0 a=recvonly m=audio 0 RTP/AVP 61 62 a=control:rtsp://cs.example.net/ode_to_joy/baseline a=rtpmap: 61 mpeg4-generic/44100 a=fmtp: 61 streamtype=5; mode=mwpp; config=""; profile-level-id=76; a=fmtp: 61 render=sasc; inline="e4"; compr=none; a=rtpmap 62 ulpfec/44100 a=fmtp: 62 rtsp://cs.example.net/ode_to_joy/fec Figure A.2 -- MWPP session description with generic FEC We refer the reader to [12] for detailed ulpfec implementation guidance. Here, we note that ulpfec supports partial packet reconstruction. This feature reduces the bandwidth of the FEC stream, but limits receivers to reconstructing only the first N octets of a lost packet. In MWPP sessions, this feature may permit receivers to reconstruct the RTP header and MIDI command section, but not the recovery journal section, of a lost MWPP packet. However, a receiver may not always be able to use a packet that does not contain a recovery journal. In particular, the recovery journal is a vital part of an MWPP packet that ends a loss event, as the receiver uses the journal to restore the integrity of the MIDI stream. To close this section, we briefly describe a second way to replace lost packets in an MWPP session. RTP defines an active tool for packet replacement, called packet retransmission [21]. In one version of packet retransmission, a receiver reports packet losses to the sender, using special RTCP receiver reports. In reply, senders supply replacement packets to the receiver, using an ancillary stream. See [21] for implementation guidance for packet retransmission systems. Lazzaro/Wawrzynek [Page 44] INTERNET-DRAFT 15 January 2003 Appendix B. Multi-party MWPP Sessions The interactive application described in the main text supports two- party sessions. In this Appendix, we modify the application to support sessions with more than two participants. We refer to these sessions as multi-party sessions. In a multi-party session, a party receives an RTP stream from each of the other parties. The application we describe uses a multicast group address to carry these RTP streams. If the session does not have access to a multicast network, the application simulates multicast with a mesh of unicast flows. In this Appendix, we show how to set up multi-party sessions, for simulated (Appendix B.1) and true (Appendix B.2) multicast scenarios. We describe sender (Appendix B.3) and receiver (Appendix B.4) modifications for multi-party sessions, and discuss scaling issues for sessions with a large number of participants (Appendix B.5). Readers should also consult Appendix A of [2] for a more detailed review of RTP and RTCP algorithms for multicast sessions. B.1 Multi-party MWPP: Session Management (simulated multicast) In this section, we describe how to set up sessions that simulate multicast transport with unicast flows. We begin with an explanation of simulated multicast (rather than true multicast) to more clearly show the link to the two-party unicast example described in the main text. Appendix B.2 covers true multicast session setup. In a simulated multicast session with N parties, each party sends its RTP and RTCP streams to N-1 other parties. Thus, a mesh of 2*N*(N-1) unicast flows acts to simulate a true multicast network. If certain conditions are met, N session descriptions may be used to define this N- party session, just as two session descriptions define the two-party session in the main text (Figures 1 and 2). However, in the general case, an N-party session requires N*(N-1) session descriptions. To show how N session descriptions may define an N-party session, we consider how the N-th party joins an existing N-1 party session. The N- th participant prepares a session description, which specifies the unicast network address and port on which it accepts RTP streams. The N- th party distributes this session description to the other parties, perhaps using a SIP conference server [14]. In return, the N-th party receives a copy of the session description for each of the other N-1 parties. In total, N unique session descriptions define the session. Lazzaro/Wawrzynek [Page 45] INTERNET-DRAFT 15 January 2003 The SDP render parameter (Appendix C.5 of [1]) may be used to define the rendering method for the session. If so, all session descriptions specify the same MIDI renderer. For example, the renderer may specify a library of SAOL instrument models (Appendix C.5.1 of [1]). A party selects timbres from the library in-band, by sending MIDI Program Change (0xB) commands. We now discuss software implementation issues for simulated multicast sessions. At the start of a session, the application chooses its synchronized source ID (ssrc), using the method described in [2]. The application also prepares a single pair of socket descriptors (rtp_fd and rtcp_fd) that it uses to send and receive MWPP streams. Figures 3 and 4 in the main text shows code for initializing rtp_fd and rtcp_fd. Whenever the application receives a session description from a new party, it creates an address_info structure (Figure B.1) for the party. The application initializes the rtp_addr and rtcp_addr address_info fields to match the RTP and RTCP destination addresses coded in the session description. Figure 6 in the main text shows code for initializing rtp_addr and rtcp_addr. The application maintains a list of active address_info structures. To send an RTP packet, the application sends a copy of the packet to the rtp_addr field of every address_info structure in the list, using the code shown in Figure 7 in the main text. The application also maintains a second data structure about each active party, the party_info structure (Figure B.2). Unlike address_info, a new party_info structure is not created in response to a new session description. Instead, party_info structures are created dynamically in response to the RTP stream, as we describe below. The primary identifier for a party_info structure is the synchronized source ID (ssrc) for this party. The ssrc is encoded in the header of RTP and RTCP packets sent by the party. The application stores active party_info structures in a hash table that is indexed by the ssrc. At the start of a session, this table is empty. typedef struct address_info { struct sockaddr_in * rtp_addr; /* where to send RTP stream */ struct sockaddr_in * rtcp_addr; /* where to send RTCP stream */ } address_info; Figure B.1 -- Addresses receivers maintain for an active party Lazzaro/Wawrzynek [Page 46] INTERNET-DRAFT 15 January 2003 typedef unsigned long uint32; /* must be 4 octets */ typedef struct party_info { uint32 ssrc; /* SSRC (synchronized source ID) */ char * cname; /* RTCP canonical name */ /* How well does this party is receive our stream? */ uint32 last_ehsnr; /* most recent EHSNR (Appendix B.3), or 0 */ /* How well we are receiving this party's stream? */ uint32 hi_seq_ext; /* highest received RTP seqnum (extended) */ struct jrec_stream * jrecv[CSYS_MIDI_NUMCHAN]; /* see Figure 10 */ /* other party-specific state (such as RTCP statistics) not shown */ } party_info; Figure B.2 -- State receivers maintain on an active stream To receive RTP packets, the application checks the rtp_fd socket descriptor, using the code shown in Figure 5 of the main text. If an RTP packet has arrived, the application examines the SSRC field of the RTP header, and uses it to locate the party_info structure for the packet in the hash table. If a party_info structure is not found, the application creates a new party_info structure and adds it to the table. The ssrc variable is set to the SSRC header value, and the hi_seq_ext field is set to the RTP sequence number header value. The new party is considered to be on probation, until future RTP packets indicate correct RTP behavior (see Appendix A of [2]). Once a new party is added to a session, the application performs session management tasks for the party, as described in Section 3 of the main text. The application transmits an RTCP stream to the parties in the address_info list, using rtcp_fd to send RTCP packets to each rtcp_addr address. The application also accepts RTCP packets on rtcp_fd, and uses the RTCP SSRC header field to locate the correct party_info structure for the packet. If no party_info exists for the SSRC value, the application creates a party_info structure for the party, following the methods described earlier in the section. Lazzaro/Wawrzynek [Page 47] INTERNET-DRAFT 15 January 2003 The application also checks to see if a party has changed its ssrc value. RTCP SDES CNAME (canonical name) packets are used to perform this check, as described in [2]. Note that the cname string in party_info codes the canonical name (username@address) of each party. The application also checks for the exit of the party, as signalled by RTCP BYE packets, session management transactions (such as SIP BYE methods), or other means. If the party leaves a session, its party_info structure is removed from the hash table, and the rendering of its MIDI stream is gracefully ended. The deletion of the address_info structure for a departed party is a more complex issue. If a session management transaction ends the session, the transaction contains information to identify the address_info structure to delete. Otherwise, the application must associate the ssrc value of party with its network address. In many cases, the source address of RTP and RTCP packets from the party correlates with the fields of the address_info structure. In other situations, data in the origin line (o=) of the session description correlates with the canonical name of the party (stored in the party_info cname variable). Finally, we note that the association between party_info and address_info structures may be of use throughout the session. For example, RTCP receiver packets may be unicast to the interested party instead of multicast to all parties, improving efficiency. Note that this optimization affects the calculation of the RTCP transmission interval [2]. B.2 Multi-party MWPP: Session Management (true multicast) In this section, we describe how to set up a multi-party session that uses a multicast group address. We refer to these sessions as true multicast sessions. A single session description is sufficient to define a true multicast session that hosts an arbitrary number of parties. The format of the session description is similar to the format of a two-party session description (Figure 1 in the main text), except that the connection (c=) line defines a multicast group address. [5] describes the SDP syntax for multicast group addresses in detail. Lazzaro/Wawrzynek [Page 48] INTERNET-DRAFT 15 January 2003 The session description may use the SDP render parameter (Appendix C.5 of [1]) to define the rendering method for the session. If so, all parties use this MIDI rendering method. For example, the renderer may specify a library of SAOL instrument models (Appendix C.5.1 of [1]). A party selects timbres from the library in-band, by sending MIDI Program Change (0xB) commands. We now discuss software implementation issues for true multicast sessions. Because multicast coding techniques vary by operating system, we do not include code fragments in this section. At the start of a session, the application chooses its synchronized source ID (ssrc), using the method described in [2]. The application also prepares a single pair of socket descriptors (rtp_fd and rtcp_fd) that it uses to send and receive MWPP streams. When an application joins a session (perhaps via a SIP conference server [14]), it receives a session description. The application prepares data structures to code the RTP (rtp_addr) and RTCP (rtcp_addr) multicast group address and port information defined in the session description. Once address preparation is complete, the application starts sending its RTP stream on rtp_fd, using rtp_addr. As in the simulated multicast case, the application maintains a party_info structure (Figure B.2) for each party in the session. The discussion of party_info hash table management in Appendix B.1 also holds for true multicast sessions. In an ongoing session, the application performs the session management tasks described in Section 3 of the main text. The discussion of these tasks for simulated multicast sessions (Appendix B.1) also holds for true multicast sessions. Finally, in some situations, an application may require a custom SDP render parameter for each sender. In this case, N true multicast session descriptions are necessary for an N party session. Each session description defines the same multicast group address. B.3 Multi-party MWPP: Sender Issues In this section, we modify the sender implementation described in Sections 4 and 5 to support multi-party sessions. We modify the recovery journal trimming algorithm (Section 5.4) to handle RTCP receiver reports from several parties. We also discuss how senders handle parties that join a session mid-stream. Apart from these issues, the sender described in the main text is compatible with multi-party sessions. Lazzaro/Wawrzynek [Page 49] INTERNET-DRAFT 15 January 2003 Section 5.4 describes an algorithm for trimming the Recovery Journal Sending Structure (RJSS) encoding of the checkpoint history. This algorithm assumes a single receiver listens to the sent stream. To trim the RJSS, the sender examines the RTCP receiver reports from the receiver, and extracts the extended highest sequence number (EHSNR) field from the report. The sender adjusts the EHSNR to reflect its own sequence number prefix, and uses the adjusted EHSNR to trim irrelevant data from the RJSS. This trimming algorithm relies on the following observation: if the EHSNR indicates that a packet with sequence number K has been received, MIDI commands sent in packets with sequence numbers I <= K may be removed from the RJSS without violating the recovery journal mandate defined in Section 4 of [1]. This observation does not hold for multi-party sessions, as several receivers may be listening to the stream. We modify this observation to be valid for multi-party sessions, in the following way. We examine the most recent ENSHR values reported by each receiver, and determine the lowest adjusted ENSHR value. If this value indicates that a packet with sequence number K has been received, MIDI commands sent in packets with sequence numbers I <= K may be removed from the RJSS without violating the recovery journal mandate defined in [1]. We now describe a multi-party RJSS trimming algorithm, that is based on the above observation. When a sender receives an RTCP receiver report, it determines the EHSNR coded by the report, using the algorithm described in Section 5.4. The sender also extracts the SSRC field of the receiver report, and locates the party_info structure (Figure B.2) associated with the ssrc. If the adjusted EHSNR matches the last adjusted EHSNR value received for this party (stored in last_ehsnr in party_info), the algorithm ends. Otherwise, last_ehsnr in party_info is updated with the adjusted EHSNR. If the first RTCP receiver report has not yet arrived for a new party, the RJSS may not be trimmed, and the algorithm ends. Otherwise, the sender loops through all party_info structures, and locates the lowest last_ehsnr value. The sender uses this last_ehsnr value to trim the RJSS, using the procedure described in Section 5.4. Note that for multi-party sessions that use a true multicast network, senders may not be aware that a new party has joined the session until the first RTP or RTCP packet has arrived from the party, or until a session management tool notifies the sender of the new party. During this interval, the sender is not able to satisfy the recovery journal mandate (Section 4 of [1]) for the new party. We discuss precautions receivers should take during this interval in Section B.4. Lazzaro/Wawrzynek [Page 50] INTERNET-DRAFT 15 January 2003 In addition, when a new party joins a session, the party needs to become aware of the current state of the MIDI streams it receives. For example, a Control Change (0xB) command for the channel volume controller (0x07) may have been sent on a stream before the new party joined the session, and the checkpoint history of the stream may no longer contain the command. Senders may bring a new receiver up to date in several ways, depending on the type of multi-party session. For simulated multicast sessions, senders may temporarily add state to the recovery journal for the benefit of a new party. Senders add this state to the journal when it becomes aware of a new party, and remove this state once it receives an RTCP receiver report from the party. This method works because a receiver is obligated to parse the recovery journal of the first RTP packet received, and a sender in a simulated- multicast session is able to ensure that this first packet contains the temporary state. However, in a true multicast session a new party may accept its first RTP packet from a sender before the sender is aware of the new party. In this case, the new party may never parse the temporary state data encoded in the recovery journal for its benefit. To solve this problem, senders insert commands into the MIDI command stream to inform a new party of the current state of the stream. If the new party has already joined the session, the new party sees the state data in the command stream. If the new party joins the session late, the new party sees the state data in the checkpoint history coded in the recovery journal. B.4 Multi-party MWPP: Receiver Issues In this section, we describe receiver implementation issues for multi- party sessions. Upon receipt of an RTP packet, the receiver uses the RTP SSRC header field to locate the party_info structure for the stream. The party_info structure contains state variables for the received stream (jrecv[] and hi_seq_ext) that are used in the receiver algorithms in Sections 6 and 7 in the main text. In most respects, these two-party algorithms are compatible with multi-party sessions. However, one multi-party incompatibility does occur in true multicast sessions. When a new party joins a multicast group, the party may begin processing RTP packets from senders that are not aware that the new party is listening. If packet loss occurs on these streams, the recovery journal of the packet that ends the loss event may not cover the loss experienced by the new party. This problem occurs because the sender has not yet seen an RTCP receiver report from the new party. Lazzaro/Wawrzynek [Page 51] INTERNET-DRAFT 15 January 2003 To handle this issue, a new party in a true multicast session should handle packet loss events with caution, until it has positive evidence that senders are aware of its presence. This evidence may take the form of RTCP receiver reports from the sender concerning the RTP stream sent by the new party. Alternatively, this evidence may take the form of information from a session management tool. Before such evidence is available, the new party should implement the precautions described in Appendix D.3, the Appendix that shows how receivers implement MWPP without RTCP. B.5 Multi-party MWPP: Scaling Issues In this section, we discuss how multi-party MWPP sessions scale to large numbers of participants. We consider sender and receiver scaling separately, as some sessions may consist of a small number of send/receive parties and a large number of receive-only parties. The memory requirements for MWPP multi-party senders scale linearly with the number of receivers listening to the stream. For a true multicast implementation, a sender uses 8 octets per listener (the ssrc and last_ehsnr party_info fields in Figure B.2). A simulated multicast implementation uses more memory per listener, to store the network addresses for each party (the rtp_addr and rtcp_addr addr_info fields in Figure B.2). Sender processor requirements also scale linearly with the number of listeners, for the simple sender algorithms we describe in this Appendix. Linear algorithms occur in several places. For example, the multi-party recovery journal trimming algorithm (Appendix B.3) uses a linear search to locate the lowest outstanding last_ehsnr value. As a second example, the all-to-all mesh method of multicast simulation (Appendix B.1) in inherently linear with the number of parties. If sub- linear scaling is required, these algorithms may be replaced with more efficient alternatives. Another type of sender scaling concerns the RTP stream bandwidth. Stream bandwidth is not constant with the number of receivers of the stream, because the operation of the recovery journal trimming algorithm is affected by number of receivers. As described in [2], a receiver sends RTCP receiver reports at a slower rate as the number of session participants increases. As a result, the recovery journal is trimmed at a slower rate, and so the RTP bandwidth increases. The rate of bandwidth growth depends on the nature of the MIDI stream, as we discuss in Appendix A.4 of [13]. Lazzaro/Wawrzynek [Page 52] INTERNET-DRAFT 15 January 2003 Next, we discuss how MWPP receivers scale to handle a large number of senders. True multicast sessions with a large number of parties may find use in experimental musical performances and multi-player games. The dominant type of receiver scaling is MIDI renderer scaling. As the number of senders increase, so do the number of simultaneous notes that sound. However, most MIDI rendering systems fail in a graceful way when asked to render too many MIDI notes at once. For example, older notes may be ended prematurely, or new note events may be selectively dropped depending on a priority scheme. Network-related processing also scales with the number of notes, not the number of senders. The memory requirements for MWPP multi-party receivers scale linearly with the number of senders. The jrecv[] array in party_info, which holds state for recovery journal processing, dominates the memory footprint. In practice, an implementation may be able to significantly reduce the size of jrecv[], by using the SDP parameters defined in Appendix C.1 to restrict the use of the journal. Lazzaro/Wawrzynek [Page 53] INTERNET-DRAFT 15 January 2003 Appendix C. MWPP and Reliable Transport RTP and MWPP may be carried over a variety of transport protocols. In the main text, we describe how to send MWPP streams over UDP, an unreliable transport protocol. In this Appendix, we describe how to send MWPP streams over reliable byte-stream protocols (such as TCP). An MWPP application chooses a transport type based on several factors. Factors that favor using UDP include: o UDP may exhibit lower latency than TCP. If the packet loss on a network is non-negligible, head-of-line blocking degrades TCP latency performance relative to UDP. However, if a network link is nearly lossless, UDP and TCP exhibit the same latency. o Multi-party MWPP applications may choose UDP in order to use multicast transport (Appendix B.2). MWPP applications that use TCP are limited to simulating multicast (Appendix B.1). o UDP may be the only transport option in low-cost embedded environments. Factors that favor using reliable transport (such as TCP) include: o TCP MWPP applications may omit recovery journal support, thus saving development costs, because TCP is reliable. UDP MWPP requires recovery journal support when used on lossy networks. o TCP may be able to pass through network middleboxes (such as firewalls). These middleboxes sometimes block all UDP traffic. o The reliable nature of TCP supports archival applications. In an archival application, the receiver intends to save an exact record of the MIDI stream to long-term storage. If TCP latency issues are a concern, applications may send two copies of the stream: a UDP copy for real-time monitoring and a TCP copy for archiving. A two-stream approach may be simpler than adding archival support to a UDP stream via retransmission [21]. In this Appendix, we describe how to use TCP in two-party interactive applications. We discuss how to define session descriptions that specify TCP streams (Appendix C.1) and describe methods for sending and receiving RTP streams (Appendix C.2). These two-party methods may also be useful in multi-party sessions that simulate multicast with a mesh of unicast flows (Appendix B.1). Lazzaro/Wawrzynek [Page 54] INTERNET-DRAFT 15 January 2003 The interactive TCP techniques in Appendices C.1-2 may also be applied to content streaming. However, the Real Time Streaming Protocol (RTSP), a session management tool for streaming media, supports interleaving RTP and RTCP streams in the RTSP TCP control stream. We discuss interleaving MWPP over RTSP in Appendix C.3. C.1 MWPP over TCP: Session Management In this section, we show how to set up two-party interactive sessions that use TCP. As in two-party UDP example in the main text, two session descriptions define a two-party TCP session. Figures C.1 and C.2 show these session descriptions. These descriptions use the methods defined in [9] and [10] for TCP session setup. The session descriptions indicate that the first party (Figure C.1) intends to initiate a TCP connection to port 16112 of 192.0.2.94 for the RTP stream, and a TCP connection to port 16113 of 192.0.2.94 for the RTCP stream. The second party intends to accept these TCP connections. Once the connections are established, RTP and RTCP streams flow bidirectionally through them. v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=Example t=0 0 c=IN IP4 192.0.2.105 m=audio 9 TCP RTP/AVP 101 a=rtpmap: 101 mwpp/44100 a=direction:active Figure C.1 -- TCP session description for first participant. v=0 o=second 2520644554 2838152170 IN IP4 second.example.net s=Example t=0 0 c=IN IP4 192.0.2.94 m=audio 16112 TCP RTP/AVP 96 a=rtpmap: 96 mwpp/44100 a=direction:passive Figure C.2 -- TCP session description for second participant. Lazzaro/Wawrzynek [Page 55] INTERNET-DRAFT 15 January 2003 Note that the direction attribute (a=) lines code the role each party plays in establishing a connection. Also note that the connection (c=) and media (m=) lines in TCP sessions indicate the network address and port on which a party accepts connections. A party that does not accept connections places the discard port (9) in its media line. This session description example shows one approach to setting up a two- party TCP session. See [9] [10] for enhancements and alternatives to this example, that may be more robust in the presence of network middleboxes (such as firewalls and network address translators). We now discuss software implementation issues for the session described in Figures C.1 and C.2. Unlike the UDP example in the main text, session setup for this TCP example is asymmetrical, as the first and second parties play different roles in the TCP connections. The first party initializes SOCK_STREAM socket descriptors for the RTP (rtp_fd) and RTCP (rtcp_fd) sessions, using the socket() system call. The first party also prepares sockaddr_in structures for the RTP (192.0.2.94:16112) and RTCP (192.0.2.94:16113) network destinations for the second party. Then, the first party establishes the RTP and RTCP connections, using the connect() system call. Once the connections are established, RTP and RTCP streams flow in both directions over rtp_fd and rtcp_fd. The second party initializes SOCK_STREAM socket descriptions for the RTP (rtp_init) and RTCP (rtcp_init) sessions, using the socket() command, and binds these sockets to the RTP (192.0.2.94:16112) and RTCP (192.0.2.94:16113) addresses listed in its own session description (Figure C.2). Then, the second party calls listen() on rtp_init and rtcp_init, and awaits connections from the first party. Once the first party responds, the second party uses accept() to assign the RTP stream to the rtp_fd socket descriptor, and the RTCP stream to the rtcp_fd socket descriptor. Once this assignment occurs, RTP and RTCP streams flow in both directions over rtp_fd and rtcp_fd. Once a session begins, the parties exchange RTCP traffic over rtcp_fd. Each RTCP packet is preceded by a two-octet unsigned integer value, sent in network byte order (big-endian), that specifies the number of octets in the RTCP packet that follows. This framing method follows [10]. Apart from RTCP packet framing, the parties perform session housekeeping duties using the methods described in Section 3 of the main text. Lazzaro/Wawrzynek [Page 56] INTERNET-DRAFT 15 January 2003 C.2 MWPP over TCP: Sending and Receiving In this section, we describe how parties send and receive MWPP RTP packets over TCP connections. Each RTP packet sent over a TCP connection is preceded by a two-octet unsigned integer value, in network byte order, that declares the number of octets in the RTCP packet that follows. This framing method follows [10]. By default, TCP sessions do not use the journalling system. The j_sec parameter overrides this default (Section C.1.1 of [1]). In the default no-journal case, senders have a responsibility to send an RTP packet stream in sequence-number order, without packet loss or reordering, and receivers may assume a perfect packet stream. In this default case, the recovery journal sending (Section 5) and receiving (Section 7) algorithms described in the main text are irrelevant. However, the core sending and receiving algorithms described in Sections 4 and 6 are quite relevant to TCP sessions, as these algorithms center on latency, bandwidth, and congestion issues. Although TCP provides congestion control, interactive performance may benefit if TCP senders use the congestion control methods described in Section 4.3. Senders may use the interarrival jitter field [2] of RTCP receiver reports to sense network congestion. Finally, we note that if the j_sec parameter configures a TCP stream to use the recovery journal, the RTP packet stream is not guaranteed to arrive in sequence-number order. This mode of operation may be used to tunnel a UDP MWPP stream through a network barrier via TCP. In this case, senders and receivers implement the recovery journal system as described in Sections 5 and 7. C.3 MWPP over TCP: RTSP Interleaving The TCP methods used in the interactive example in Appendices C.1-2 may also be used in content streaming applications. However, the Real Time Streaming Protocol (RTSP, [6]) provides more convenient methods for TCP streaming. In normal RTSP usage, a receiver (or in RTSP terminology, a client) may contact an RTSP server to engage in session control transactions. Using RTSP, a client may request a session description for a media stream, whose format is coded with the Session Description Protocol. However, these session descriptions usually omit session transport details (network addresses, ports, etc). Lazzaro/Wawrzynek [Page 57] INTERNET-DRAFT 15 January 2003 Instead, RTSP clients and servers exchange transport details in the Transport header lines of RTSP methods and responses. In most cases, the net result of the transaction is identical to what would happen if the transport information was carried in the session description. Based on the Transport header line data, servers and clients set up unicast or multicast flows, and RTP and RTCP streams are carried on these flows. However, this observation is not true for RTSP interleaved mode. Interleaved mode, signalled by the use of the interleaved parameter on the RTSP transport line, does not result in the creation or use of separate transport connections for media. Instead, RTP and RTSP packets are interleaved onto the TCP stream that carries the RTSP control transactions. This method has the advantage that the media blocking by network middleboxes is not possible, because the RTSP TCP connection already exists. Section 10.13 of [6] describes RTSP interleaving in detail, including packet framing methods. Here, we note one MWPP RTSP issue (as normatively specified in [1]). If an MWPP stream is interleaved over a TCP RTSP control stream, by default the MWPP payload does not use the recovery journal. In this case, the server has the duty to send an RTP packet stream in sequence-number order, without packet loss or reordering, and client may assume a perfect packet stream. To set up an interleaved MWPP stream that uses the recovery journal, the j_sec parameter (as defined in Section C.1.1 of [1]) must be present on the RTSP Transport line of the SETUP method response (if the server supports journalling mode) and perhaps the request (if the client is requesting journalling mode). If an interleaved MWPP stream uses the recovery journal, the server is not obliged to send an RTP stream free from packet loss or reordering events, and the client must be prepared to handle these events. Lazzaro/Wawrzynek [Page 58] INTERNET-DRAFT 15 January 2003 Appendix D. Using MWPP without RTCP MWPP works best with RTCP. MWPP implementations use RTCP in several ways. MWPP senders use RTCP receiver reports as a feedback signal for congestion control (Section 4.3). MWPP senders also use receiver reports to trim the checkpoint history of the recovery journal (Section 5.4). MWPP receivers use RTCP sender reports for multi-stream synchronization (Appendix E). However, MWPP does not require RTCP, and session descriptions may specify MWPP streams that do not use RTCP. Embedded devices may choose to only support MWPP without RTCP, to reduce memory requirements. In this Appendix, we describe how MWPP senders and receivers perform checkpoint history management, congestion control, and playback synchronization without the use of RTCP. In the sections below, we modify the application described in Sections 2-7 to work without RTCP. D.1 MWPP without RTCP: Session Management Section 2 shows how to start a two-party interactive session that uses RTCP. Figures 1 and 2 show the session descriptions that define the session. Figures D.1 and D.2 show modified versions of these session descriptions, that specify a session without RTCP. Figure D.1 defines how the first party wishes to receive its stream; Figure D.2 defines how the second party wishes to receive its stream. The modified session descriptions disable RTCP by using bandwidth (b=) lines to set the RTCP session bandwidth to zero [8]. The session descriptions also use the SDP j_update parameter to define checkpoint management policies (Appendix C.1.1 of [1]) that are suitable for sessions that do not use RTCP. In the modified session descriptions, the first party (Figure D.1) receives a stream that uses the anchor policy for checkpoint management. In the anchor policy, the checkpoint packet identity is fixed for the entire session. This policy works well for streams that use a few MIDI command types (Appendix A.4 of [13]). The ch_unused parameter specifies the MIDI commands that the sender does not use (Appendix C.1.2 in [1]) The second party (Figure D.2) receives a stream that uses the open-loop policy for checkpoint management. In this policy, the sender updates the checkpoint packet at regular intervals, dropping older commands from the checkpoint history. After a packet loss, receivers determine if the checkpoint history covers the loss event, by using the checkpoint sequence number coded in the recovery journal header (Figure 7 in [1]). If the loss is not covered, the receiver executes MIDI commands to restore the integrity of the stream. Lazzaro/Wawrzynek [Page 59] INTERNET-DRAFT 15 January 2003 Note that for certain MIDI command types, receivers are not able to recover from an uncovered loss event. For example, if a Control Change (0xB) command for the channel volume controller (0x07) is prematurely dropped from the checkpoint history, a receiver has no way to ascertain the correct volume. To address this issue, the modified session descriptions use the ch_anchor parameter to protect fragile chapters (Appendix C.1.2 in [1]). Open-loop senders never drop ch_anchor chapters from the checkpoint history. The session setup algorithms defined in Figures 3-7 may be used for sessions that do not use RTCP, by deleting the code referencing the RTCP socket descriptor rtcp_fd and the RTCP address rtcp_addr. Once the session begins, session housekeeping tasks are identical to those described in Section 3, except that tasks related to RTCP are not performed. v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=Example t=0 0 c=IN IP4 192.0.2.94 m=audio 16112 RTP/AVP 96 a=rtpmap: 96 mwpp/44100 b=RS:0 b=RR:0 a=fmtp: 101 j_update=anchor; ch_unused=ATMDVQEX; Figure D.1 -- Session description for first participant. v=0 o=second 2520644554 2838152170 IN IP4 second.example.net s=Example t=0 0 c=IN IP4 192.0.2.105 m=audio 5004 RTP/AVP 101 a=rtpmap: 101 mwpp/44100 b=RS:0 b=RR:0 a=fmtp: 101 j_update=open-loop; ch_unused=ATMDVQEX; ch_anchor=PWC; Figure D.2 -- Session description for second participant. Lazzaro/Wawrzynek [Page 60] INTERNET-DRAFT 15 January 2003 D.2 MWPP without RTCP: Sender Issues In this section, we modify the sender implementation described in Sections 4 and 5 to support sessions that do not use RTCP. The modifications affect congestion control, multi-stream synchronization, and checkpoint history management. Senders use RTCP receiver reports as feedback signals for congestion control (Section 4.3). If RTCP is not in use, other congestion measures may be available. For example, session management may take place via peer-to-peer UDP SIP [14] transactions. In this case, the loss rate of SIP response or ACK packets measures the combined congestion of the forward and reverse paths. Note that this method is inferior to RTCP receiver reports in several ways: SIP transactions may occur infrequently and SIP proxies in the network path may degrade the loss data. RTCP also plays a role in multi-stream synchronization. RTCP sender reports link the RTP timestamp clock to an absolute time clock. Receivers may use this absolute reference to synchronize multiple MWPP streams. However, MWPP also supports multi-stream synchronization without RTCP, using the SDP zerosync parameter (Appendix C.4.2 of [1]). The zerosync parameter defines sender behaviors for RTP timestamp generation, as we discuss in detail in Appendix E. Finally, sender modifications are necessary for checkpoint history management. As the two parties defined in Figure D.1 and D.2 use different checkpoint management policies, we describe two separate modifications of the sender implementation. The first party (Figure D.1) uses the anchored checkpoint management policy. In this policy, the checkpoint packet identity is fixed for the entire session. To implement this policy, we delete the trimming algorithm described in Section 5.4 from the sender. The second party (Figure D.2) uses the open-loop policy for checkpoint management. In this policy, the sender updates the checkpoint packet at regular intervals, dropping older commands from the checkpoint history. To implement this policy, we replace the RTCP-oriented trimming algorithm described in Section 5.4. The new algorithm implements a control system to maintain the RTP stream bandwidth below a pre-defined limit. Whenever the stream bandwidth exceeds the limit, the sender reduces the size of the recovery journal. To perform a trimming operation, the sender reduces the size of the checkpoint history of the journal. The sender trims or deletes journal chapters to match the shortened history, and updates the recovery Lazzaro/Wawrzynek [Page 61] INTERNET-DRAFT 15 January 2003 journal header to code the new checkpoint packet. Chapters protected by the ch_anchor parameter are never trimmed. This algorithm is conservative, in that it aims to approximate the behavior of the anchored checkpoint algorithm, subject to bandwidth limits. Other open-loop trimming algorithms are possible. For example, a trimming algorithm may aim to minimize the bandwidth of the stream, for a given level of protection against uncovered packet loss. D.3 MWPP without RTCP: Receiver Issues In this section, we modify the receiver described in Sections 6 and 7 to support sessions that do not use RTCP. The modifications affect algorithms for multi-stream synchronization and packet loss recovery. Receivers use RTCP sender reports for multi-stream synchronization. However, MWPP also supports multi-stream synchronization without RTCP, using the SDP zerosync parameter (Appendix C.4.2 of [1]). The zerosync parameter codes the relative timing of MWPP streams in a session. Appendix E describes how receivers may use relative timing information to synchronize multiple MWPP streams. The packet loss recovery algorithm described Section 7 is incompatible with the open-loop checkpoint management policy (Appendix C.1.1 of [1]). The open-loop policy is often used in sessions that do not use RTCP, such as the session defined in Figure D.2. We now describe a modified version of the loss recovery algorithm, that fixes this incompatibility. In the modified algorithm, upon the detection of a loss event, the receiver compares the checkpoint packet sequence number coded in the recovery journal header to the highest RTP sequence number previously seen in the stream. This comparison is performed modulo 2^16, and uses standard methods (described in [2]) to avoid tracking errors during rollover. If the checkpoint packet sequence number is less than the highest RTP sequence number, the recovery journal may not code complete recovery information for the packet loss event. We refer to this condition as an uncovered loss event. When an uncovered loss occurs, the chapter-specific recovery algorithms use a modified recovery strategy, that takes the incomplete nature of the chapter data into account. Note that for the session defined in Figure D.2, only Chapter N is vulnerable to uncovered losses, as the ch_anchor parameter protects Chapters W, P, and C, and the ch_unused parameter excludes all other chapters from the journal. Lazzaro/Wawrzynek [Page 62] INTERNET-DRAFT 15 January 2003 In the case of an uncovered loss event, the Chapter N recovery procedure described in Section 7.2 performs in the following way. If data for a note number appears in Chapter N, the algorithms described in Section 7.2 are executed as normal. If data for a note number does not appear in Chapter N, and the vel[] array in jrec_chaptern indicates the note is currently on, we assume a NoteOff command was lost for the note number. Thus, we execute the recovery procedure that would occur if the bit for note number in the journal NoteOff bit array is set, as described in Section 7.2. Lazzaro/Wawrzynek [Page 63] INTERNET-DRAFT 15 January 2003 Appendix E. Multi-stream MWPP Sessions The session descriptions shown in the main text (Figure 1 and 2 in Section 2) use one media (m=) line. This media line specifies the delivery of a single MIDI name space (16 voice channels + systems). In this Appendix, we show how to use several MWPP media lines in a session description, and discuss why applications may wish to do so. We use the term multi-stream session to describe sessions of this type. Multi-stream session descriptions may use grouping lines [11] to specify the synchronized playback of media streams. For example, a session description may group an audio and a video media stream, to specify a lip-synced presentation. In a similar manner, grouping lines may serve to synchronize MIDI flows in multi-stream MWPP sessions. Multi-stream MWPP session descriptions may also specify: o Name space relationships. The SDP MWPP parameter midiport may be used to code name space relationships between MWPP streams. For example, one stream may code voice channels 1-8 of a MIDI name space, and a second stream may code voice channels 9-16 of the same MIDI name space. o Synchronization mechanics. MWPP streams code discrete events, not a continuous media flow. The MWPP SDP parameter zerosync codes information to improve synchronization lock-in time for event streams. In this Appendix, we describe common uses for multi-stream MWPP sessions (Appendix E.1). We also discuss MWPP sender and receiver synchronization (Appendix E.2) and name space (Appendix E.3) issues. E.1 Multi-Stream Session Scenarios In this section, we describe several applications of multi-stream sessions, and show session descriptions for each application. We also discuss how applications set up multi-stream sessions. A simple form of multi-stream session specifies two or more independent MWPP streams. If MWPP streams are independent, the MIDI name spaces of the streams are unrelated, and the streams are rendered independently. However, grouping lines may be used to specify synchronized rendering of the streams (the LS grouping semantics [11]). Grouping lines may also be used to specify that several MWPP streams represent the same data flow (the FID grouping semantics [11]), and thus are copies of each other. Lazzaro/Wawrzynek [Page 64] INTERNET-DRAFT 15 January 2003 Figure E.1 and E.2 show an independent multi-stream session. One party (Figure E.1) is a musician, who has two keyboard controllers in her rig. The session description maps each keyboard to a separate MWPP stream, and uses the sendonly attribute to code that the controllers do not accept MIDI input. The NTP timestamps coded in the RTCP sender reports for the two streams share a common clock source. v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=Two keyboards driving independent synths. t=0 0 c=IN IP4 192.0.2.105 a=sendonly a=group: LS top bottom m=audio 5004 RTP/AVP 96 i=Keyboard top a=rtpmap: 96 mwpp/44100 a=mid:top a=fmtp: 96 zerosync=18293 m=audio 5006 RTP/AVP 96 i=Keyboard bottom a=rtpmap: 96 mwpp/44100 a=mid:bottom a=fmtp: 96 zerosync=23893 Figure E.1 -- Player definition for the two-keyboard example. v=0 o=second 102902938 9837465 IN IP4 second.example.net s=Two keyboards driving independent synths. t=0 0 c=IN IP4 192.0.2.94 a=recvonly a=group: LS top bottom m=audio 16112 RTP/AVP 96 i=Synth for top keyboard a=rtpmap: 96 mwpp/44100 a=mid:top m=audio 16114 RTP/AVP 96 i=Synth for bot keyboard a=rtpmap: 96 mwpp/44100 a=mid:bottom Figure E.2 -- Synth definition for the two-keyboard example. Lazzaro/Wawrzynek [Page 65] INTERNET-DRAFT 15 January 2003 The second party (Figure E.2) is a computer that runs two separate music synthesizers, one for each stream. The recvonly attribute in the session description of the second party codes that the synths do not send MIDI back to the keyboards. The first party sends its MWPP streams to the network addresses defined in the session description of the second party. In these session descriptions, the LS grouping parameter codes that the streams are to be synchronized on playback. Apart from this synchronization, rendering proceeds independently for the two streams. To aid synchronization lock at the start of the session, the session description for the musician uses the MWPP SDP parameter zerosync. In Appendix E.2, we describe how senders and receivers use data coded by the zerosync parameter. Figures E.3 and E.4 show another independent multi-stream session. In this example, the first party (Figure E.3) is a musician with a single keyboard controller in her rig. Identical copies of the MIDI commands from the controller are sent over both streams. The grouping parameter FID codes that both streams represent the same flow of data. v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=One keyboard sent to a synth and a recorder t=0 0 c=IN IP4 192.0.2.105 a=sendonly a=group: FID synth recorder m=audio 5004 RTP/AVP 96 i=The synth stream a=rtpmap: 96 mwpp/44100 a=mid:synth m=audio 9 TCP RTP/AVP 96 i=The recorder stream a=rtpmap: 96 mwpp/44100 a=mid:recorder a=direction:active Figure E.3 -- Player definition for the one-keyboard example. Lazzaro/Wawrzynek [Page 66] INTERNET-DRAFT 15 January 2003 v=0 o=second 2520644554 2838152170 IN IP4 second.example.net s=One keyboard sent to a synth and a recorder t=0 0 a=recvonly a=group: FID synth recorder m=audio 16112 RTP/AVP 96 i=The synth stream c=IN IP4 192.0.2.94 a=rtpmap: 96 mwpp/44100 a=mid:synth m=audio 16114 TCP RTP/AVP 96 i=The recorder stream c=IN IP4 192.0.2.21 a=rtpmap: 96 mwpp/44100 a=mid:recorder a=direction:passive Figure E.4 -- Target definition for the one-keyboard example. The second party (Figure E.4) renders the first MWPP stream into audio, and records the second MWPP stream to disk. The rendered stream uses UDP transport, for lowest latency. The archived stream uses TCP transport, to ensure an accurate recording. As the two applications run on different machines, the two streams have different network addresses (192.0.2.94 and 192.0.2.21). We now show examples of sessions do not use independent streams. Instead, the streams in the session are in a relationship (Appendix C.4 of [1]). Streams in a relationship share a common MIDI name space. For some MWPP renderers (such as sasc, defined in Appendix C.5 of [1]), the streams in a relationship also share the same renderer instance (a property defined in Appendix C.4.1. of [1]). One type of MWPP relationship is the identity relationship. All streams in an identity relationship target the same MIDI name space (16 voice channels + systems). Two MWPP streams share an identity relationship if the same value is assigned to the midiport parameter in each stream description (Appendix C.4.1 of [1]). Lazzaro/Wawrzynek [Page 67] INTERNET-DRAFT 15 January 2003 v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=Keyboard and librarian controlling a synth. t=0 0 c=IN IP4 192.0.2.105 a=group: LS keyboard librarian m=audio 5004 RTP/AVP 96 i=The keyboard stream a=rtpmap: 96 mwpp/44100 a=mid:keyboard a=sendonly a=fmtp 96 midiport=12; zerosync = 0; a=fmtp 96 ch_unused=X; m=audio 9 TCP RTP/AVP 96 i=The librarian stream a=rtpmap: 96 mwpp/44100 a=direction:active a=mid:librarian a=fmtp 96 midiport=12; zerosync = 0; a=fmtp 96 ch_unused=PWNATCMDVQE; a=sendrecv Figure E.5 -- Player for the keyboard/librarian example. v=0 o=second 2520644554 2838152170 IN IP4 second.example.net s=Keyboard and librarian controlling a synth. t=0 0 c=IN IP4 192.0.2.94 a=group: LS keyboard librarian m=audio 16112 RTP/AVP 96 i=The keyboard stream a=rtpmap: 96 mwpp/44100 a=mid:keyboard a=recvonly a=fmtp 96 midiport=12; m=audio 16114 TCP RTP/AVP 96 i=The librarian stream a=rtpmap: 96 mwpp/44100 a=mid:librarian a=direction:passive a=fmtp 96 midiport=12; a=fmtp 96 ch_unused=PWNATCMDVQE; a=sendrecv Figure E.6. -- Synth for the keyboard/librarian example. Lazzaro/Wawrzynek [Page 68] INTERNET-DRAFT 15 January 2003 Figure E.5 and E.6 define a session that contain an identity relationship. In this example, the first party (Figure E.5) is a musician, and the second party (Figure E.6) is a synthesizer. The musician uses a MIDI controller keyboard to play the synthesizer. Occasionally, the musician also uses a MIDI librarian program to update the timbre memory of the synthesizer. The librarian generates and receives MIDI System Exclusive commands, that coordinate large data transfers between the librarian and the synthesizer. The session uses two streams, in order to optimize the network transport for each type of data. The controller keyboard stream uses UDP transport, for good latency performance. The librarian stream uses TCP transport, for reliable transfer of timbre data. The streams share an identity relationship, so that both the keyboard and the librarian may issue commands in the MIDI name space of the synthesizer. The midiport parameter in each stream establishes the identity relationship. The ch_unused parameter in the librarian stream specifies that the stream uses only MIDI System Exclusive commands. The streams are synchronized using the LS grouping semantics and the zerosync parameter, to ensure correct temporal ordering of MIDI events in the two streams. The librarian stream uses the sendrecv attribute, as the MIDI handshaking protocols used by the librarian generate bidirectional traffic. Our final example uses a session with an ordered relationship. Ordered relationships accommodate applications that group MIDI streams into an extended name space (32 voice channels, 48 voice channels, etc). The first party (Figure E.7) is a sequencer program, that is playing back a 32-channel MIDI performance. Channels 1-16 are sent on the first stream, channels 17-32 are sent on the second stream. The adjacent midiport values (5 and 6) for the streams establish the ordering. The second party (Figure E.8) is an MPEG 4 Structured Audio [7] renderer (one of the sasc renderers). Structured Audio supports the concept of an extended MIDI channel number space. As normatively specified in Appendix C.4.1 of [1], both streams in this ordered relationship target a single instance of this sasc renderer. We now discuss session management issues for multi-stream sessions. At the network layer, each stream in the session is an independent entity. Multi-stream applications implement the session setup and management algorithms described in Sections 2 and 3 for each stream. Each stream uses a unique pair of network ports, accessed by separate instances of the rtp_fd and rtcp_fd socket descriptors (Sections 2). Lazzaro/Wawrzynek [Page 69] INTERNET-DRAFT 15 January 2003 v=0 o=first 2520644554 2838152170 IN IP4 first.example.net s=32-channel sequencer driving a Structured Audio renderer t=0 0 c=IN IP4 192.0.2.64 a=group: LS upper lower m=audio 5004 RTP/AVP 61 a=rtpmap: 61 mpeg4-generic/44100 a=fmtp: 61 streamtype=5; mode=mwpp; config=""; profile-level-id=74; a=fmtp: 61 midiport=5;zerosync=0; a=mid:lower a=sendonly m=audio 5006 RTP/AVP 62 a=rtpmap: 62 mpeg4-generic/44100 a=fmtp: 62 streamtype=5; mode=mwpp; config=""; profile-level-id=74; a=fmtp: 62 midiport=6;zerosync=0; a=fmtp: 62 render=sasc; url="http://www.example.com/cardinal.sasc"; a=fmtp: 62 cid="azsldkaslkdjqpwojdkmsldkfpe"; a=mid:upper a=sendonly Figure E.7 -- Sequencer for the Structured Audio example. v=0 o=second 2520644554 2838152170 IN IP4 second.example.net s=32-channel sequencer driving a Structured Audio renderer t=0 0 c=IN IP4 192.0.2.69 a=group: LS upper lower m=audio 10000 RTP/AVP 61 a=rtpmap: 61 mpeg4-generic/44100 a=fmtp: 61 streamtype=5; mode=mwpp; config=""; profile-level-id=74; a=fmtp: 61 midiport=5;zerosync=0; a=mid:lower a=recvonly m=audio 10002 RTP/AVP 64 a=rtpmap: 62 mpeg4-generic/44100 a=fmtp: 62 streamtype=5; mode=mwpp; config=""; profile-level-id=74; a=fmtp: 62 midiport=6;zerosync=0; a=fmtp: 62 render=sasc; url="http://www.example.com/cardinal.sasc"; a=fmtp: 62 cid="azsldkaslkdjqpwojdkmsldkfpe"; a=mid:upper a=recvonly Figure E.8. -- Renderer for the Structured Audio example. Lazzaro/Wawrzynek [Page 70] INTERNET-DRAFT 15 January 2003 E.2 Synchronization Issues In this section, we discuss how senders and receivers synchronize multiple MWPP streams. We begin with a review of RTP multi-stream synchronization methods. We describe how to apply these methods to MWPP streams, and discuss synchronization issues that are unique to MWPP. A common RTP synchronization task is lip-syncing audio and video streams. In a typical situation, a session begins with a multimedia server sending audio and video RTP streams to a receiver. The first packets sent in the audio and the video streams represent the same moment in media time. The receiver assumes the first packet for both streams arrived without loss, and begins to buffer the two streams. Once the buffers are sufficiently full, audio and video playback begins. As the session progresses, drift between the four clocks in the system (audio capture, video capture, audio playback, video playback) results in the audio and video falling out of sync. The receiver is unable to use the RTP timestamps of the two streams to restore sync, because these timestamps embody the capture drift. Even determining the numerical relationship between the audio and video timestamps is not trivial, as each RTP stream uses a random timestamp offset [2]. To solve this problem, the receiver uses the RTCP sender reports of the two streams. These reports are sent at frequent intervals (for a two- party session that follows the guidelines in [2], about once every 5 seconds). Each sender report codes a 64-bit Network Time Protocol wall- clock timestamp together with its associated RTP timestamp for the stream. By mapping the RTP timestamps for each stream to NTP absolute time, the receiver is able to sync the audio and video stream to the accuracy limits of the NTP clock sources. We now consider how to apply these synchronization methods to multi- stream MWPP sessions. We begin by considering MWPP receivers that use a playout buffer, such as the content streaming applications described in Appendix A. Some types of interactive receivers also use playout buffers (Section 6 of the main text). For these receivers, the standard RTP synchronizations work well once RTCP reports for all streams have arrived. However, before initial mappings of RTP timestamps to NTP timestamps for the streams are delivered by RTCP, the receiver may render the MWPP streams wildly out of sync. Lazzaro/Wawrzynek [Page 71] INTERNET-DRAFT 15 January 2003 Stream startup is a problem because the first packet in an MWPP stream codes the first MIDI command in the stream. Unless the first MIDI command in all MWPP streams in a session happen at the same moment in time, playout buffers that render the first packet of all streams simultaneously will not produced a synchronized output. The MWPP zerosync parameter, defined in Appendix C.4.1 of [1], codes start up timing information in the session descriptions. Receivers may use this information to synchronize the start up multiple MWPP streams. As normatively stated in [1], if the MWPP streams in a LS grouping use the zerosync parameter, the srate values for all streams MUST be identical. The zerosync parameter may be used in two different ways. One use of zerosync, shown in the session descriptions in Figures E.1 and E.2, uses the zerosync parameter to encode the RTP timestamp offset for each stream (in the example, 18293 for the top stream and 23893 for the bottom stream). By subtracting (modulo 2^32) the stream offset from the RTP timestamp for each packet, the receiver may recover common stream timestamps for use during the startup period. Note that if the session description transport occurs in a secure manner, this use of zerosync does not degrade RTP security. A second use of zerosync, shown in the session descriptions in Figures E.5 and E.6, sets the zerosync parameter for each stream to the special value of 0. In this use of zerosync, all MWPP streams in the session whose zerosync value is zero use the same RTP timestamp offset, and so the RTP timestamps of the streams may be directly compared. This use of zerosync weakens the security of an encrypted RTP stream, and should be avoided in secure sessions. Finally, we consider MWPP applications that do not use a playout buffer, such as the simpler interactive receiver designs described in Section 6 of the main text. In these applications, the temporal integrity of the performance is based on the assumption that the underlying network has low nominal jitter. These methods rely on careful implementations of sender and receiver algorithms, to minimize the introduction of processing jitter at the endpoints. Multi-stream versions of this architecture must be doubly careful in this regard, to avoid adding temporal offsets or jitter across streams. Receivers that do not use a playout buffer use RTP timestamps to identify packets that arrive late, as described in Section 6.1 of the main text. However, these algorithms use timestamps in a differential way, and so multi-stream versions of these receivers do not need to synchronize the timestamps of the streams. Lazzaro/Wawrzynek [Page 72] INTERNET-DRAFT 15 January 2003 E.3 Name Space Issues In this Appendix, we discuss implementation issues for streams that target the same MIDI name space (16 voice channels + systems). Streams that target the same MIDI name space share an identity relationship. Figures E.5 and E.6 of Appendix E.1 show an example session that includes an identity relationship. We focus on identity relationships because MWPP streams that target the same MIDI name space run the risk of integrity loss. The other type of stream relationship, the ordered relationship (Figures E.7 and E.8 of Appendix E.1), poses no unusual implementation issues. Identity relationships are prone to integrity loss, because an arbitrary partitioning of a MIDI name space between several streams may circumvent the recovery journal system. For example, consider an identity relationship that placed all NoteOn (0x9) commands on one stream and all NoteOff (0x8) commands on a second stream. If the two streams are not perfectly synchronized, the NoteOff pattern may slip ahead of the NoteOn pattern, and stuck notes may occur. Packet losses are also problematic in this partitioning scheme, as the recovery journal mechanism for MIDI Note commands (Chapter N, as defined in Appendix A.4 in [1]) assumes that all MIDI note commands for a channel are present in the stream. Reference [1] describes the multi-stream integrity issue, and defines normative guidelines to prevent it from occurring (Appendix C.4.1 in [1]). We summarize these guidelines below: o Session participants MUST choose a MIDI name space partitioning that does not result in rendered performances that contain indefinite artifacts. o If an artifact-free performance requires a specific temporal sequencing of commands across streams, senders MUST guarantee this sequencing. o Receivers MUST maintain the structural integrity of the MIDI name space as it merges incoming streams. This requirement includes transaction-oriented MIDI commands, such as the Registered and Non-Registered Parameter MIDI Control Change (0xB) commands. In this case, receivers assume that a transaction occurs within a single stream. The safest way to partition a MIDI name space is to place all commands affecting a voice channel, including System Exclusive commands that are associated with the channel, into one stream. In addition, Systems Lazzaro/Wawrzynek [Page 73] INTERNET-DRAFT 15 January 2003 commands with related functionality, such as the MIDI sequencer commands, should also be grouped together in a stream. However, applications requirements may conflict with these simple partitioning rules, and a more nuanced approach may be required. For example, a player may wish to route two MIDI controllers, such as a keyboard controller (generating Note commands) and a continuous controller (generating Control Change (0xB) commands), to the same synthesizer. In situations of this nature, it is safe to split one MIDI voice channel between streams that share a MIDI name space. The MIDI librarian session in Figures 5 and 6 in Appendix E.1 shows another example of nuanced multi-stream partitioning. In this session, bulk-data System Exclusive commands related to a voice channel are sent on a separate stream from interactive voice channel commands. This stream partitioning optimizes the network transport type for real-time (sent on a UDP stream) and bulk-data (sent on a TCP stream) MIDI commands. This librarian example shows the rationale for the sender and receiver responsibilities for multi-stream systems defined in [1]. Senders are responsible for correct intra-stream sequencing, because (in this example) careless sender RTP timestamps may place real-time MIDI commands on the wrong side of a bulk-data transfer. Likewise, a careless receiver implementation that did not respect MIDI merging semantics might attempt to interleave commands from the real-time stream into an ongoing bulk-data download. Lazzaro/Wawrzynek [Page 74] INTERNET-DRAFT 15 January 2003 Appendix F. References F.1 Normative References [1] John Lazzaro and John Wawrzynek. The MIDI Wire Protocol Packetization (MWPP). draft-ietf-avt-mwpp-midi-rtp-05.txt. [2] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson. RTP: A transport protocol for real-time applications. Work in progress, draft-ietf-avt-rtp-new-11.txt. [3] H. Schulzrinne and S. Casner. RTP Profile for Audio and Video Conferences with Minimal Control. Work in progress, draft-ietf-avt-profile-new-12.txt. [4] MIDI Manufacturers Association. The complete MIDI 1.0 detailed specification, 1996. http://www.midi.org [5] M. Handley, V. Jacobson and C. Perkins. SDP: Session Description Protocol. Work in progress, draft-ietf-mmusic-sdp-new-10.txt. [6] H. Schulzrinne, A. Rao, and R. Lanphier. Real Time Streaming Protocol (RTSP). Work in progress, draft-ietf-mmusic-rfc2326bis-00.txt. [7] International Standards Organization. ISO 14496 MPEG-4, Part 3 (Audio) Subpart 5 (Structured Audio) 1999. [8] S. Casner. SDP Bandwidth Modifiers for RTCP Bandwidth. draft-ietf-avt-rtcp-bw-05.txt [9] D. Yon. Connection-Oriented Media Transport in SDP. . [10] The forthcoming I-D to bring back the old RTP RFC TCP framing method. If a new framing method is chosen by the WG instead, the text attached to this reference will change to describe the new framing method. [11] G. Camarillo, G. Eriksson, J. Holler, H. Schulzrinne. Grouping of Media Lines in the Session Description Protocol (SDP). RFC 3388. [12] A. Li, F. Liu, J. Villasenor, J.H. Park, D.S. Park, Y.L. Lee, J. Rosenberg, and H. Shulzrinne. An RTP Payload Format for Generic FEC with Uneven Level Protection. draft-ietf-avt-ulp-07.txt. Lazzaro/Wawrzynek [Page 75] INTERNET-DRAFT 15 January 2003 F.2 Informative References [13] John Lazzaro and John Wawrzynek. A Case for Network Musical Performance. The 11th International Workshop on Network and Operating Systems Support for Digital Audio and Video (NOSSDAV 2001) June 25-26, 2001, Port Jefferson, New York. http://www.cs.berkeley.edu/~lazzaro/sa/pubs/pdf/nossdav01.pdf [14] J. Rosenberg, H. Schulzrinne, G. Camarillo, A. Johnston, J. Peterson, R. Sparks, M. Handley, and E. Schooler. SIP: Session Initiation Protocol. Internet Engineering Task Force, RFC 3261. [15] J. Rosenberg, R. Mahy, and S. Sen. NAT and Firewall Scenarios and Solutions for SIP. draft-ietf-sipping-nat-scenarios-00.txt. [16] Baugher, McGrew, Oran, Blom, Carrara, Naslund, and Norrman. The Secure Real-time Transport Protocol. Work in progress, draft-ietf-avt-srtp-05.txt. [17] Dominique Fober, Yann Orlarey, Stephane Letz. Real Time Musical Events Streaming over Internet. Proceedings of the International Conference on WEB Delivering of Music 2001, pages 147-154 http://www.grame.fr/~fober/RTESP-Wedel.pdf [18] C. Bormann et al. Robust Header Compression (ROHC). Internet Engineering Task Force, RFC 3095. Also see related work at http://www.ietf.org/html.charters/rohc-charter.html. [19] Sfront source code release, includes a Linux networking client that implements the MIDI RTP packetization. http://www.cs.berkeley.edu/~lazzaro/sa/ [20] The SoundWire group, http://ccrma-www.stanford.edu/groups/soundwire/ [21] Joerg Ott, Uni Bremen, Stephan Wenger, Noriyuki Sato, Carsten Burmeister, and Jose Rey. Extended RTP Profile for RTCP-based Feedback (RTP/AVPF). draft-ietf-avt-rtcp-feedback-04.txt. Lazzaro/Wawrzynek [Page 76] INTERNET-DRAFT 15 January 2003 Appendix G. Author Addresses John Lazzaro (corresponding author) UC Berkeley CS Division 315 Soda Hall Berkeley CA 94720-1776 Email: lazzaro@cs.berkeley.edu John Wawrzynek UC Berkeley CS Division 631 Soda Hall Berkeley CA 94720-1776 Email: johnw@cs.berkeley.edu Lazzaro/Wawrzynek [Page 77]