INTERNET-DRAFT Stephan Wenger draft-wenger-avt-rtcp-feedback-02.txt TU Berlin Joerg Ott Universitaet Bremen TZI 2 March, 2001 Expires September 2001 RTCP-based Feedback: Concepts and Message Timing Rules Status of this Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of RFC 2026. 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/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract Real-time media streams are not resilient against packet losses. RTP [1] provides all the necessary mechanisms to restore ordering and timing to properly reproduce a media stream at the recipient. RTP also provides continuous feedback about the overall reception quality from all receivers -- thereby allowing the sender(s) in the mid-term (in the order of several seconds to minutes) to adapt their coding scheme and transmission behavior to the observed network QoS. However, except for a few payload specific mechanisms [2], RTP makes no provision for timely feedback that would allow a sender to repair the media stream immediately: through retransmissions, retro-active FEC, or media-specific mechanisms such as reference picture selection. This document specifies a modification to the algorithm for scheduling RTCP packets in order to allow occasional timely feedback to events observed by a receiver (such a lost packets). The message format for RTCP-based feedback is defined in a companion document [7]. Wenger/Ott Expires September 2001 [Page 1] Internet Draft 24 November, 2000 1. Introduction Real-time media streams are not resilient against packet losses. RTP [1] provides all the necessary mechanisms to restore ordering and timing present at the sender to properly reproduce a media stream at a recipient. RTP also provides continuous feedback about the overall reception quality from all receivers -- thereby allowing the sender(s) in the mid-term (in the order of several seconds to minutes) to adapt their coding scheme and transmission behavior to the observed network QoS. However, except for a few payload specific mechanisms [2], RTP makes no provision for timely feedback that would allow a sender to repair the media stream immediately: through retransmissions, retro-active FEC, or media-specific mechanisms such as reference picture selection. Current mechanisms available with RTP to improve error resilience include audio redundancy coding [3], video redundancy coding [4], RTP-level FEC [5], and general considerations on more robust media streams transmission [6]. Particularly in small groups, however, virtually all kinds of all types of real-time media streams could benefit from a mechanism that would enable a sender to perform media stream repair -- including but not limited to audio, video, DTMF, and text chat streams. In some case of networks with acceptable round- trip times but scarce bandwidth, occasional retransmissions may be much preferred over continuous transmission of redundant information. For example, predictive video coding is not loss resilient. Any loss of coded data leads to annoying artifacts not only in the reproduced picture in which the loss occurred, but also in subsequent pictures. Error resilience can be achieved by spending bits to convey redundant information using source coding based mechanisms or transport based mechanisms. This can be done without the use of any feedback between the decoder(s) and the encoder. Similar consideration apply to protecting e.g. DTMF (and other tones) carried in an RTP stream [9]. Alternatively, where applicable, receivers can inform the sender through a feedback channel about a loss situation, and the sender can react accordingly. This approach provides better media quality and is more efficient with respect to the bandwidth used by the sender to achieve a given media quality. However, using feedback mechanisms is limited to certain application scenarios identified by encoder characteristics, delay constraints, and/or the number of recipients. This memo specifies a profile based upon [1] and [10] with enhanced rules for sending receiver reports to support feedback transmission reflecting the need for very low delay for conveying feedback, which is necessary to make them efficient (or workable at all). Immediate Feedback messages (FB messages) and Early Receiver Reports (Early RRs) and algorithms are specified that allow for low delay in small multicast groups, but prevent network flooding in larger ones. Special consideration is given to point-to-point scenarios. Wenger/Ott Expires September 2001 [Page 2] Internet Draft 24 November, 2000 In addition, this memo gives some consideration to specific application scenarios are the respective feedback requirements, at the moment focusing on predictive video coding. A companion document [7] discusses various types of general purpose feedback information (also allowing for extensions specific to certain media payload) and defines an RTCP packet format to transmit FBs in an RTP environment. It can be used in conjunction with all payload specifications for predictive video coding schemes currently available for RTP. 2. Motivation 2.1 Example: Predictive Video Coding 2.1.1 Video Encoder-decoder synchronicity Most current video coding schemes for compressed video, such as the ITU-T H.261 and H.263 and ISO/IEC MPEG[124] employ a mechanism known as Inter Picture Prediction. Each picture is divided into macroblocks of uniform size. For each macroblock, one or more motion vectors may be identified and transmitted. The residual signal after motion compensation is DCT-transformed, quantized, entropy coded, and transmitted as well. The encoder reconstructs, based on this information, a so-called reference picture, which is used to perform the motion compensation and residual signal coding steps for the subsequent picture. Since the reference picture is generated using only such information that is also available at the decoder, the reference picture is identical to the reconstructed picture at the decoder. Having identical reference pictures at the encoder and decoder is referred to as encoder-decoder-synchronicity. Whenever data is damaged or lost on the way between the encoder and the decoder, the reconstructed picture at the decoder is no more identical with the encoder's reference picture -- the encoder-decoder synchronicity is lost. Any loss of the encoder-decoder synchronicity results in annoying artifacts at the decoder. Because the prediction of subsequent pictures in the decoder is based on a damaged reference picture, the annoying artifacts are present not only in the picture in which the loss occurred; they propagate to all subsequent pictures, until, through source coding based mechanisms, the encoder-decoder synchronicity is restored. Therefore, the goal of systems employing predictive video coding in a lossy environment must be to keep the encoder-decoder synchronicity, or, if this is not possible, to regain that synchronicity as quickly as possible. 2.1.2. Non-feedback based mechanisms Avoiding the loss of the encoder-decoder synchronicity corresponds to avoiding the loss of coded picture data. Such a task can be performed on the transport layer. In RTP environments, the use of packet-based FEC is a good example for such a technique. (The use of Wenger/Ott Expires September 2001 [Page 3] Internet Draft 24 November, 2000 TCP or reliable multicast as the transport for media streams would be an even better one but is inappropriate for low-delay (interactive) real-time systems.) FEC schemes, interleaving, and other means for repairing real-time media streams may also add additional delay and significant bit rate overhead without being able to guarantee compensation of virtually all packet losses. Once the encoder-decoder synchronicity is lost, only source coding oriented mechanisms can help to regain it. One common way is to send a non-predictively coded picture (known as Intra picture). Intra pictures have the disadvantage of being several times bigger than predictively coded pictures (Inter pictures). Therefore, sending Intra pictures has negative implications both on the bandwidth and (in bandwidth limited environments) delay. Another way is to use Intra macroblock refresh. Here, certain parts of the picture (those affected by a packet loss) are coded non-predictively in order to resynchronize the encoder and decoder over time. Intra macroblock refresh has better delay characteristics then full Intra pictures because the picture size can be kept constant, but is less efficient in terms of bit rate/distortion than full Intra pictures. More sophisticated means such as Reference Picture Selection (RPS) are also available in modern video coding standards. Systems not employing feedback channels may use any combination of the mechanisms described above to add error resilience -- at the cost of added bit rate and, sometimes, added delay. The number of additional bits spent for error resilience can be adapted using the long-term packet loss rate information in the RTCP receiver reports. But, even when using such adaptive means, it is still likely that systems spend many more bits then theoretically necessary to achieve error resilience in order to be on the safe side. Plus, as regular RTCP feedback is aimed at longer terms, reactivity to sudden losses is limited. In all practical applications today this means that fewer bits are available for non redundant picture data, and hence the overall picture quality suffers. 2.1.3 Feedback based systems Feedback-based systems try to avoid spending too many bits for redundant information by informing the encoder about a loss situation at the decoder(s). The encoder can then react accordingly and spend redundant bits only when needed possibly only for the part of the picture that was effected by the loss -- thereby reducing the number of redundant bits and leaving more bits for useful information. As a result, a higher reproduced picture quality can generally be expected when feedback channels are available. Similar to the observations of section 2.1.2, transport and source coding based mechanisms can be distinguished that react on loss situations reported by feedback. Transport based systems employing feedback react media unaware, by re-transmitting lost packets. TCP is a good example for a protocol Wenger/Ott Expires September 2001 [Page 4] Internet Draft 24 November, 2000 following such a scheme. Transport-based feedback in real-time and/or multicast environments is a complex matter and subject of a lot of engineering and research in and outside of the IETF. This specification is not concerned with pure transport-based feedback. Source coding based mechanisms may react upon the arrival of a feedback message indicating a loss situation by adding bits that restore, or at least make an effort to restore, the encoder-decoder synchronicity. This process has to be performed by a real-time encoder. However, schemes were reported, that allow the use of feedback also for non-real-time encoders by storing multiple representations of the same data (e.g. Inter and Intra coded), and dynamically switching between those representations. Several types of feedback messages, called Feedback Messages or FB messages, can be defined for such a case. An FB message can be as simple as a Boolean condition, indicating for example the loss of a full picture (and, therefore, the need of a full Intra picture transmission). Other feedback messages may contain more complex information such as information about the damage of a spatial region of the picture. A special form consists of a message the format and semantics of which are not known at the transport level, because they are defined in the video codec standards. 2.2 Feedback Messages Most FB messages contain negative acknowledge information, indicating an erroneous situation at the decoder. In others, the nature of the acknowledge (positive, negative, or both) is part of the feedback message itself. When used in multicast environments, positive acknowledge must not be used. This document assumes that feedback messages are transmitted using RTCP packets. RTCP messages from the receivers to the sender cannot be sent at any possible time, in order to prevent traffic explosion in case of large multicast groups. Instead, the bit rate for all RTCP messages of all receivers together has to obey a maximum fraction of the total RTP session bit rate, yielding a very limited bit rate budget for a single receiver when having a large multicast group. This, in turn, leads to an increased average delay when the size of the receiving multicast group grows. (see section 6 of [1] for details) This specification defines an algorithm that adheres to the bit rate limitations for the feedback channel on the long term, but allows short-term overdrafting for any receiver (but not all of them simultaneously). Thus, the algorithm allows for better real-time performance then the one specified in [1]. Traffic explosion in such cases in which many receivers identify a picture damage simultaneously is prevented by dithering. As this specification assumes a sender that has full control over its transmission bit rate (e.g. a real-time encoder), there is no scaling Wenger/Ott Expires September 2001 [Page 5] Internet Draft 24 November, 2000 problem on the forward channel. Any reaction to negative feedback generates additional bits, which have to be conveyed but this is taken from the sender's total bit rate budget. The encoder can take this into account by, for example, changing the encoding mode, packet size, and so forth. The sender is also free to simply ignore feedback messages. Adjusting the tradeoff between the reproduced media quality of all receivers of a multicast group and the amount of additional repair traffic is a media-dependent, very complex task and is not covered in this specification. Finally, frequent RTCP-based feedback messages may provide additional input to the sender(s)'s congestion control algorithms and thus improve its reactivity towards network congestion. Feedback messages as well as sender and receiver behavior are to be specified in separate documents (such as [7]). Such specifications need to consider that, frequently, packet loss is an indication of network congestion and thus define mechanisms for media-specific congestion control in the presence of feedback as defined in this memo. 2.3. Applications and Relationships to other Standards This specification is based on RTCP, which implies its use in an RTP environment. RTP itself is used in a variety of systems such as in SIP- or H.323-based multimedia conferencing/telephony, SAP-announced Mbone conferences, and RTSP-based media streaming. As for the video codecs, there is currently a small set of standards that are, for the purpose of this discussion, roughly comparable. Many mechanisms for regaining encoder-decoder synchronicity are applicable to all video codecs. Others require certain tools (such as Reference Picture Selection, aka NEWPRED) that are available only in certain versions of the standards, and/or optional tools whose use must be negotiated prior to being used. A few RTP payload specifications such as RFC 2032 [2] already define a feedback mechanism for some of the coding algorithms considered in this specification. An application capable of performing both schemes MUST use the feedback mechanism defined in this specification, although, for backward compatibility reasons, it MUST also be capable to conform to the feedback scheme defined in the respective RTP payload format, if this is required by that payload format. Also, audio, DTMF, and text streams could benefit from more immediate feedback even though the redundancy payload formats work well for these media. All kinds of non-interactive media streams (such as RTSP-controlled media streaming applications) could benefit significantly as without interactivity there is more time available for media repair. Wenger/Ott Expires September 2001 [Page 6] Internet Draft 24 November, 2000 2.4 Remarks on the size of the multicast group This specification prevents traffic explosion on the feedback channel in a very similar way as RTP does, with the exception of allowing individual receivers to overdraft their bit rate budget from time to time. This is necessary in order to allow for low delay, which is needed by the algorithms reacting to Feedback messages. This scaling, however, limits the usefulness of this mechanism in multicast groups from a certain size upwards (where the size threshold depends on a number of parameters including loss rate, frame rate, number of packets per frame, and session bandwidth). The maximum size of the multicast group is soft and also depends on application requirements and is therefore not specified here. Considerations on the multicast group sizes will be presented in section 3.5. 2.5 Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [8] 3. Low delay RTCP Feedback Two components constitute RTCP-based feedback as described in this memo: . Status reports are contained in SR/RR messages and are transmitted at regular intervals as part of compound RTCP packets (which also include SDES and possibly other messages); these status reports provide an overall indication for the recent reception quality of a media stream. RTP [1] define rules for the transmission of these status reports. . Feedback messages as defined in a companion document [7] that indicate loss or reception of particular pieces of a media stream (or provide some other form of rather immediate feedback on the data received). Rules for the transmission of feedback messages are newly introduced in this memo. As discussed in [7], RTCP Feedback (FB) messages are just another RTCP message type. Thus multiple FB messages may be combined in a single RTCP packet. FB messages may be sent in full compound RTCP packets along with SR/RR, SDES, and other RTCP messages. Or they may be transmitted in minimal compound RTCP FB packets (which only contain the RR/SR and an encryption prefix if necessary to reduce the message size). RTCP packets that do not contain FB messages are referred to as non-FB RTCP packets. Wenger/Ott Expires September 2001 [Page 7] Internet Draft 24 November, 2000 3.1 Algorithm Outline FB messages are part of the RTCP control streams and are thus subject to the same bandwidth constraints as other RTCP traffic. This means in particular that it may not be possible to report a packet loss at a receiver immediately back to the sender. However, the value of feedback given to a sender typically decreases over time -- in terms of the media quality as perceived by the user at the receiving end and/or the cost required to achieve media stream repair. RTP [1] specifies rules when compound RTCP packets should be sent. This specification modifies those rules in order to allow applications to timely report media loss or reception events, since most algorithms that use FB messages are very critical to the feedback timing. See section 5 and following for a discussion of FB messages and the impact of delay on the performance these FB types. The modified algorithm can be outlined as follows: Normally, when no FB messages have to be conveyed, compound RTCP packets are sent following the rules of RTP [1]. If a receiver detects the need for an FB message, the receiver first checks whether it has already seen a corresponding FB message from any other receiver (which it can do with all FB messages that are transmitted via multicast; for unicast sessions, there is no such delay). If this is the case then the receiver refrains from sending the FB message, and continues to follow the regular RTCP sending schedule. If the receiver has not yet seen a similar FB message from any other receiver, it checks whether it has recently exceeded its RTCP bit rate budget to transmit another FB message (without waiting for its regularly scheduled RTCP transmission time). Only if this is not the case, it sends the FB message, after waiting a short, random dithering interval period (in case of multicast). FB messages are sent as part of minimal compound RTCP packets . Full compound RTCP packet are interspersed as per [1] in regular intervals of at least five seconds. 3.2 Modes of Operation RTCP-based feedback may operate in one of three modes (figure 1): a) Immediate feedback mode: the group size is below a certain threshold (the FB threshold) which gives each receiving party sufficient bandwidth to transmit the feedback traffic for the intended purpose. This means, for each receiver there is enough bandwidth to report each event it is supposed/expected to by means of a virtually "immediate" Early RTCP packet. The group size threshold is a function of a number of parameters including (but not necessarily limited to) the type of feedback used (e.g. ACK vs. NACK), bandwidth, packet rate, packet loss probability, media type, codec, and -- again depending on the type Wenger/Ott Expires September 2001 [Page 8] Internet Draft 24 November, 2000 of FB used -- the (worst case or observed) frequency of events to report (e.g. frame received, packet lost). A special case of this is the ACK mode (where positive acknowledgements are used to confirm reception of data) which is restricted to point-to-point communications. b) In Early RTCP mode, the group size and other parameters no longer allow each receiver to react to each event that would be worth (or needed) to report. But feedback can still be given sufficiently often so that it allows the sender to adapt the media stream and thereby increase the overall reproduced media quality. c) From some group size upwards, it is no longer useful to provide feedback from individual receivers at all -- because of the time scale in which the feedback could be provided and/or because in large groups the sender(s) have no chance to react to individual feedback anymore. As the feedback algorithm described in this memo scales, there is no need for an agreement on the precise values of the respective "thresholds" within the group. Hence the borders between all these modes are fluent. ACK feedback V :<- - - - NACK feedback - - - ->// : : Immediate || : Feedback mode ||Early RTCP mode Regular RTCP mode :<=============>||<=============>//<=================> : || -+---------------||---------------//------------------> group size 2 || Application-specific FB Threshold = f(rate,loss,codec,...) Figure 1: Modes of operation The respective thresholds depend on a number of technical parameters (of the codec, the transport, the feedback used, etc.) but also on the respective application scenarios. Section 3.5 provides some useful hints (but no complete precise calculations) on estimating these thresholds. 3.3 Definitions a) Let the media stream be transmitted at a (roughly) constant packet rate f (in packets per second). This results in an average inter-packet interval of tau=1/f. Wenger/Ott Expires September 2001 [Page 9] Internet Draft 24 November, 2000 b) Let T_rtt be the maximum round trip time as measured by RTCP (if available to the receiver). Note that this may be asymmetric. d) Let t_rr and t_(rr-1) be the time for the next (last) scheduled RTCP RR transmission calculated prior to reconsideration. Let T_rr + t_(rr-1) = t_rr. (In RTP [1] these are termed tp, tn, respectively). d) Let t_e be the time for which a feedback packet is scheduled. e) Let t_dither_max be the maximum interval for which an RTCP feedback packet may be additionally delayed (to prevent implosions). f) Let T_fd be the delay for the feedback message that a certain packet P caused to return to the sender after reception of P. g) Let S be the number of active senders in the RTP session. h) Let N be the current estimate of the number of receivers in the RTP session. The feedback situation for an event to report at a receiver is depicted in figure 2 below. At time t0, such an event (e.g. a packet loss is detected at the receiver. The receiver decides -- based upon current T_rtt, group size, and other (application-specific) parameters -- that a feedback message shall be sent back to the sender. To avoid an implosion of immediate feedback packets, the receiver delays transmission of the compound feedback packet by a random amount T_fd (with the random number evenly distributed in the interval [0, T_dither_max]. Transmission of the compound RTCP packet is then scheduled for t_e = t0 + T_fd. The T_dither_max parameter is chosen based upon the group size, the RTCP bandwidth constraints, and, if available, the round-trip time. In addition, the receiver may take into account a number of other parameters (such as the estimated round-trip time, the type of feedback to be provided) to possibly extend the upper bound for the feedback while ensuring that the feedback information still will make sense when it reaches the sender. If a compound RTCP feedback packet is scheduled, the time slot for the next scheduled compound RTCP packet is updated accordingly to a new t_rr. Wenger/Ott Expires September 2001 [Page 10] Internet Draft 24 November, 2000 event to report detected | | RTCP feedback vXXXXXXXXXXXXXXXXXXXX ) ) |---+--------+-------------+-----+------------| |--------+---------> | | | | ( ( | | t0 te | t_(rr-1) t_rr \_______ ________/ \/ T_dither_max Figure 2: Event report and parameters for Early RTCP scheduling 3.4 Early RTCP Algorithm Assume an active sender S0 (out of S senders) and a number N of receivers with R being one of these receivers. Assume further that R has verified that using feedback mechanisms is reasonable at the current constellation (which is highly application specific and hence not specified in this memo). Then, the following rules apply to transmitting a Feedback Messages as minimal compound RTCP packet: Initially, R sets allow_early := TRUE. At a point in time t0, R has transmitted the last RTCP RR packet at t_(rr-1) and has scheduled the next transmission (prior to reconsideration) for t_rr. Now R detects the need to transmit a feedback message (e.g. because a media "unit" needs to be ACKed or NACKed) at time t0. R first checks whether there is still a feedback packet waiting for transmission. If so, the new feedback message is appended to the packet and the increased RTCP packet size is updated in the RTCP bandwidth calculation (which may later lead to an adjustment of t_rr); the schedule for the waiting RTCP feedback packet remains unchanged. If no feedback message is already awaiting transmission a new (minimal) compound RTCP feedback message is created and the interval T_dither_max is chosen as follows: i) If the session is a unicast session (group size = 2) then T_dither_max := 0. Wenger/Ott Expires September 2001 [Page 11] Internet Draft 24 November, 2000 ii) If the receiver has an RTT estimate to the originator of the media unit to provide feedback about, then / T_rtt/2 if T_rtt/2 > 10ms T_dither_max := < \ 10ms otherwise. iii) If the receiver does not have an RTT estimate to the originator, then / T_rr/2 if T_rr/2 < 100ms T_dither_max := < \ 100ms otherwise. (Note: These values are *still* open to discussion.) (Note that application-specific feedback considerations may make it worth while to increase T_dither_max beyond this value.) Then, R checks whether its next regularly scheduled RTCP packet is within the time bounds for the RTCP FB (t_e + T_dither_max > t_rr). If so, no Early RTCP is scheduled; instead the FB message is appended to the regular RTCP packet and the RTCP bandwidth calculation is updated to reflect the additional RTCP size. The updated bandwidth calculation may result in a slightly increased t_rr (=t_rr') but, even if t_rr' > t_e + T_dither_max, this does not change the updated transmission time t_rr'. (Q: if the FB is piggybacked onto a regularly scheduled RTCP RR message but the same or a superset of the feedback information is received from another receiver, should the FB then be removed from the compound RR/FB and its transmission time be revised again from t_rr' to t_rr as calculated before?) Otherwise, R MUST check whether it is allowed to transmit an Early RTCP packet (allow_early == TRUE). If so, R schedules an Early RTCP packet for t_e = t0 + RND * T_dither_max with the RND function evenly distributed between 0 and 1. If R receives an RTCP feedback packet (indicating the same or a superset of the feedback information R wanted to transmit) before t_e is reached, the FB information is discarded and the transmission schedule for the next RR packet is reset to t_rr as calculated before. Otherwise, when t_e is reached, R creates an RR, appends the FB information, and transmits the RTCP packet. R then sets allow_early := FALSE and recalculates t_rr := t_e + 2*T_rr. As soon as R sends its next regularly scheduled RTCP RR (at the new t_rr), it sets allow_early := TRUE again. Wenger/Ott Expires September 2001 [Page 12] Internet Draft 24 November, 2000 If allow_early == FALSE then R checks the time for the next scheduled RR: if t_rr - t0 < T_dither_max then R creates an FB message for transmission along with the RTCP packet at a then slightly modified t_rr' (see above). Otherwise, R does not send an RTCP feedback message at all. In regular RTCP intervals as specified by [1] (i.e. at most every five seconds), a full compound RTCP packet is sent (which may also contain a feedback message if one has been created according to the above rules and scheduled for transmission along the full compound RTCP message). The E bit in the message header [7] is used upon reception to detect whether this RTCP feedback message was sent as Early RTCP or not. Hence, a feedback message that is sent as an Early RTCP packet MUST set the E bit in the message header to "1". Feedback messages piggy- backed on regularly scheduled RTCP packets will MUST set the E bit to "0". 3.5 Considerations on the Group Size This section intends to give some brief guidelines to the group sizes at which the various feedback modes may be used. 3.5.1 ACK mode The group size MUST be exactly two participants, i.e. point-to-point communications. Unicast addresses SHOULD be used in the session description. For unidirectional as well as bi-directional communication between two parties, 2.5% of the RTP session bandwidth are available for feedback. Assuming a ratio of 1:10 for minimal to full compound RTCP packets, at 64kbit/s, a receiver can report 2.5 events per second back to the sender, at 256kbit/s 10 events and so forth. From 768kbit/s upwards, a receiver would be able to acknowledge each individual frame (not packet!) in a 30 fps video stream. ACK strategies have to be defined accordingly to work with these bandwidth limitations. 3.5.2 NACK mode Negative acknowledgements (or similar types of feedback) have to be used for all groups larger than two. Whether or not the use of Immediate or Early RTCP packets should be considered depends upon a number of parameters including session bandwidth, codec, special type of feedback, number of senders and receivers, among many others. Wenger/Ott Expires September 2001 [Page 13] Internet Draft 24 November, 2000 The crucial parameters -- to which all of the above can be reduced -- is the allowed minimal interval between two RTCP reports and the number of events that presumably need reporting per time interval. The minimum interval is derived from the available RTCP bandwidth and the expected average size of an RTCP packet. The number events to report e.g. per second may be derived from the packet loss rate and sender's rate of transmitting packets. From these two values, the allowable group size for the Immediate feedback mode can be calculated. The upper bound for the Early RTCP mode then solely depends on the acceptable quality degradation, i.e. how many events per time interval may go unreported. Example: If a 256kbit/s video with 30 fps is transmitted through a network with an MTU size of some 1500 bytes, then, in most cases, each frame would fit in its own packet leading to a packet rate of 30 packets per second. If 5% packet loss occurs in the network (equally distributed, no inter-dependence between receivers), then each receiver will have to report 3 packets lost each two seconds. Assuming a single sender and more then three receivers yields 3.75% of the RTCP bandwidth allocated to the receivers and thus 9.6kbit/s. Assuming further a size of 100 bytes for the average compound RTCP packet allows 12 RTCP packets to be sent per second or 24 in two seconds. If every receiver needs to report three packets, this yields a maximum group size of 8 receivers if all loss events shall be reported. The rules for transmission of immediate RTCP packets should provide sufficient flexibility for most of this reporting to occur in a timely fashion. Extending this example to determine the upper bound for Early RTCP mode leads to the following considerations: assume that the underlying coding scheme and the application (as well as the tolerant users) allow in the order of one loss without repair per two seconds. Thus the number of packets to be reported by each receiver decreases to two per two seconds second and increases the group size to 12. Assuming further that some number of packet losses are correlated, feedback traffic is further reduced and group sizes of some 15 to 20 can be reasonably well supported using Early RTCP mode. 3.6 Summary of decision steps 3.6.1 General Hints Before even considering whether or not to send RTCP feedback information an application has to determine whether this mechanism is applicable: 1) An application has to decide whether -- for the current ratio of packet rate with the associated (application-specific) maximum feedback delay and the currently observed round-trip time (if available) -- feedback mechanisms can be applied at all. Wenger/Ott Expires September 2001 [Page 14] Internet Draft 24 November, 2000 This decision may obviously be based upon (and dynamically revised following) regular RTCP reception statistics. 2) The application has to decide whether -- for a certain observed error rate, assigned bandwidth, frame rate, and group size -- (and which) feedback mechanisms can be applied. Regular RTCP provides valuable input to this step, too. 3) If these tests pass, the application has to follow the rules for transmitting Early RTCP packets or regularly scheduled RTCP packets with piggybacked feedback. 3.6.2 Session Description Attributes A number of additional SDP parameters may be used to describe a session. These are defined as session level and/or media level attributes: 3.6.1.1 RTCP Feedback a=rtcp-fb: {"ack"|"nack"|extension} params This attribute is used to indicate the feedback (to be) supported by the sender. "ack" MUST only be used if the media session is allowed to operate in ACK mode as defined in 3.6.1.2. It is up to the recipients whether or not they send feedback information and up to the sender(s) to make use of feedback provided. 3.6.1.2 Unicasting If an m= line in the SDP describing a session indicates unicast addresses for a particular media type (and does not operate in multi- unicast mode with all recipients listed explicitly but still addressed via unicast), the RTCP feedback MAY operate in ACK feedback mode. 4. Format of RTCP Feedback messages The general format of the FB messages are defined in [7]. 5. Security Considerations RTP packets transporting information with the proposed payload for mat are subject to the security considerations discussed in the RTP specification [1]. This implies that confidentiality of the media streams is achieved by encryption. If the entire stream (extension data and AU data) is to be secured and all the participants are expected to have the keys to decode the Wenger/Ott Expires September 2001 [Page 15] Internet Draft 24 November, 2000 entire stream, then the encryption is performed in the usual manner, and there is no conflict between the two operations (encapsulation and encryption). The need for a portion of stream (e.g. extension data) to be encrypted with a different key, or not to be encrypted, would require application level signaling protocols to be aware of the usage of the XT field, and to exchange keys and negotiate their usage on the media and extension data separately. 6. Acknowledgements Large parts of the syntax and the text concerned with RPS and NEWPRED were borrowed from an early I-D from Fukunaga et. al. that was concerned with MPEG-4 ES packetization. 7. Full Copyright Statement Copyright (C) The Internet Society (2001). All Rights Reserved. This document and translations of it may be copied and furnished to others, and derivative works that comment on or otherwise explain it or assist in its implementation may be prepared, copied, published and distributed, in whole or in part, without restriction of any kind, provided that the above copyright notice and this paragraph are included on all such copies and derivative works. However, this document itself may not be modified in any way, such as by removing the copyright notice or references to the Internet Soci- ety or other Internet organizations, except as needed for the purpose of developing Internet standards in which case the procedures for copyrights defined in the Internet Standards process must be fol- lowed, or as required to translate it into languages other than English. The limited permissions granted above are perpetual and will not be revoked by the Internet Society or its successors or assigns. This document and the information contained herein is provided on an "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF MER- CHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE." 8. Authors' Addresses Stephan Wenger (stewe@cs.tu-berlin.de) TU Berlin Sekr. FR 6-3 Franklinstr. 28-29 Wenger/Ott Expires September 2001 [Page 16] Internet Draft 24 November, 2000 D-10587 Berlin Germany Joerg Ott (jo@tzi.uni-bremen.de) Universitaet Bremen TZI MZH 5180 Bibliothekstr. 1 D-28359 Bremen Germany 4. Bibliography [1] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson, "RTP - A Transport Protocol for Real-time Applications," Internet Draft, draft-ietf-avt-rtp-new-08.txt, Work in Progress, July 2000. [2] T. Turletti and C. Huitema, "RTP Payload Format for H.261 Video Streams, RFC 2032, October 1996. [3] C. Perkins, I. Kouvelas, O. Hodson, V. Hardman, M. Handley, J.C. Bolot, A. Vega-Garcia, and S. Fosse-Parisis, "RTP Payload for Redundant Audio Data," RFC 2198, September 1997. [4] C. Bormann, L. Cline, G. Deisher, T. Gardos, C. Maciocco, D. Newell, J. Ott, G. Sullivan, S. Wenger, and C. Zhu, "RTP Payload Format for the 1998 Version of ITU-T Rec. H.263 Video (H.263+)," RFC 2429, October 1998. [5] C. Perkins and O. Hodson, "2354 Options for Repair of Streaming Media," RFC 2354, June 1998. [6] J. Rosenberg and H. Schulzrinne, "An RTP Payload Format for Generic Forward Error Correction,", RFC 2733, December 1999. [7] S. Fukunaga, N. Sato, K. Yano, A. Miyazaki, K. Hata, R. Hakenberg, C. Burmeister, "Low Delay RTCP Feedback Format," Internet Draft draft-fukunaga-low-delay-rtcp-02.txt, Work in Progress, February 2001. [8] S. Bradner, "Key words for use in RFCs to Indicate Requirement Levels," RFC 2119, March 1997. [9] H. Schulzrinne and S. Petrack, "RTP Payload for DTMF Digits, Telephony Tones and Telephony Signals," RFC 2833, May 2000. [10] H. Schulzrinne and S. Casner, " RTP Profile for Audio and Video Conferences with Minimal Control," Internet Draft draft-ietf- avt-profile-new-09.txt, July 2000. Wenger/Ott Expires September 2001 [Page 17] Internet Draft 24 November, 2000 Appendix A: Considerations On Video This section of this memo covers feedback messages for a Picture Loss Indication (PLI), Slice Loss Indication (SLI), and Reference Picture Selection Indication (RPSI). PLI indicates the loss of a full picture and roughly corresponds to the Fast Intra Request known from H.320 systems and from RFC 2032 (H261 packetization). Algorithms using SLI can be found under the acronym Automatic Repeat Request (ARQ) in the signal processing literature. Reference Picture Selection, aka NEWPRED, is available in certain profiles of MPEG-4 (version 2 and later) and as an optional mode in H.263 (version 2 and later). The packet format specified in this document is open to extensions so that future feedback mechanisms can easily be integrated. All these messages use the payload specific feedback format as defined in [7], using PT=PSFB and the FMT field to further distinguish between the three subtypes. These messages are defined for payload types indicating H.263 and MPEG-4. Note that the Bit 00 of the first (counting from 1) 32-bit word in the messages described below is placed in Bit 08 of the fourth (counting from 1) 32-bit word of the payload type specific feedback message. A.1 Message Type 1: Picture Loss Indication (PLI) A.1.1 Semantics With the Picture Loss Indication message a decoder informs the encoder about the loss of one or more full pictures A.1.2 Format PLI does not require parameters. Therefore, the length field MUST be 0, and there MUST NOT be Feedback Control Information. A.1.3 Timing Rules The timing follows the rules outlined in section 3. In systems that employ both PLI and other FB types it may be advisable to follow the regular RTCP RR timing rules, since PLI is not as delay critical as other FB types. A.1.4 Remarks PLI messages typically trigger the sending of full Intra pictures. Intra Pictures are several times larger then predicted (Inter) pictures. Their size is independent of the time they are generated. In most environments, especially when employing bandwidth-limited links, the use of an Intra picture implies an allowed delay that is a significant multitude of the typical frame duration. An example: If the sending frame rate is 10 fps, and an Intra picture is assumed to be 10 times as big as an Inter picture (not an unrealistic Wenger/Ott Expires September 2001 [Page 18] Internet Draft 24 November, 2000 assumption, see [] for details), then a full second of latency has to be accepted. In such an environment there is no need for a particular short delay in sending the feedback message. Hence waiting for the next possible time slot allowed by RFC1889bis RTCP timing rules does not negatively influence system performance. A.2 Message Type 2: Slice Lost Indication A.2.1 Semantics With the Slice Lost Indication a decoder can inform an encoder that it was unable to decode one, or several consecutive, macroblocks. The encoder can take appropriate action in order to re-synchronize encoder and decoder by means of its choice, typically by sending the lost macroblocks in Intra mode. This feedback message SHALL NOT be used for video codecs with non-uniform, dynamically changeable macroblock sizes such as H.263 with enabled Annex Q. In such a case, an encoder cannot always identify the corrupted spatial region. A.2.2 Format When FBT indicates a Slice Lost Indication, then there is one additional UCI field the content of which is in the following format: 0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | First | Number | TR | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ First: 13 bits The macroblock (MB) address of the first lost macroblock. The MB numbering is done such that the macroblock in the upper left corner of the picture is considered macroblock number 1 and the number for each macroblock increases from left to right and then from top to bottom in raster-scan order (such that if there is a total of N macroblocks in a picture, the bottom right macroblock is considered macroblock number N). Number: 13 bits The number of lost macroblocks, in scan order as discussed above. TR: 6 bits The six least significant bits of the Temporal Reference of the picture. A.2.3 Timing Rules The efficiency of algorithms using the Slice Lost Indication is reduced greatly when the Indication is not transmitted in a timely fashion. Motion compensation propagates corrupted pixels that are not reported as being corrupted. Therefore, the use of the algorithm discussed in section 3 is highly recommended. Wenger/Ott Expires September 2001 [Page 19] Internet Draft 24 November, 2000 Constraints on T_dither_max to be discussed. A.2.4 Remarks The First field of the UCI defines the first macroblock of a picture as 1 and not, as one could suspect, as 0. This was done to align this specification with the comparable mechanism available in H.245. The maximum number of macroblocks in a picture (2**13 or 8192) corresponds to the maximum picture sizes of the ITU-T and ISO/IEC video codecs. If future video codecs offer larger picture sizes and/or smaller macroblock sizes, then an additional feedback message has to be defined. The six least significant bits of the Temporal Reference field are deemed to be sufficient to indicate the picture in which the loss occurred. Algorithms were reported that keep track of the regions effected by motion compensation, in order to allow for a transmission of Intra macroblocks to all those areas, regardless of the timing of the FB [TBP.]. While, when those algorithms are used, the timing of the FB is less critical then without, it has to be observed that those algorithms correct large parts of the picture and, therefore, have to transmit many for bits in case of delayed FBs. A.3 Message Type 3: Reference Picture Selection Indication A.3.1 Semantics Modern video coding standards such as MPEG-4 visual version 2 or H.263 version 2 allow the use of older reference pictures then the most recent one. Typically, a first-in-first-out queue of reference pictures is maintained. If an encoder has learned about a loss of encoder-decoder synchronicity, a known-as-correct reference picture can be used. As this reference picture is temporally further away then usual, the resulting predictively coded picture will use more bits. Both MPEG-4 and H.263 define a binary format for the _payload_ of an RPSI message that includes information such as the temporal ID of the damaged picture and the size of the damaged region. This bit string is typically small _- a couple of dozen bits -_, of variable length, and self-contained, i.e. contains all information that is necessary to perform reference picture selection. Note that both MPEG-4 and H.263 allow the use of RPSI with positive feedback information as well. That is, all corrected pictures are reported. Any form of positive feedback MUST NOT be used when in a multicast environment (reporting positive feedback about individual reference pictures at RTCP intervals is not expected to be of much use anyway). For point-to-point communication, positive feedback MAY be used but, again, the bit rate budget of RTCP feedback will prevent the use in most scenarios anyway. A.3.2 Format Wenger/Ott Expires September 2001 [Page 20] Internet Draft 24 November, 2000 When FB indicates an RPSI, then the length field is set to the number of bits of the following bit string that contains the RPS information. This bit string follows byte aligned in the UCI field. Bit padding is used to achieve 32-bit word alignment of the UCI message (and the whole packet). A.3.3 Timing Rules RPS is even more critical to delay then algorithms using SLI. This is due to the fact that the older the RPS message is, the more bits the encoder has to spend to achieve encoder-decoder synchronicity. See [TBP.] for some information about the overhead of RPS for certain bit rate/frame rate/loss rate scenarios. Therefore, RPS messages should typically be sent as soon as possible, employing the algorithm of section 3. Constraints on T_dither_max to be discussed. A.3.4 Remarks TBD. Wenger/Ott Expires September 2001 [Page 21]