Internet Engineering Task Force Omar Elloumi INTERNET DRAFT Stefaan De Cnodder Kenny Pauwels Alcatel July, 1999 Expires January, 2000 Usefullness of three drop precedences in Assured Forwarding service Status of this Memo This document is an Internet-Draft and is in full conformance with 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/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract This informational memo points out advantages of using three drop precedences (compared to two) in the Assured Forwarding (AF) Per- Hop-Behavior (PHB) when the traffic is composed of responsive TCP and non responsive UDP traffic. Our simulation results show that while two drop precedences can, in some cases, guarantee to an aggregate consisting of TCP traffic the full utilization of the reserved bandwidth, they cannot allow to access the unreserved bandwidth in the presence of high bit rate unresponsive UDP traffic. When three drop precedences are appropriately used TCP aggregates can fully use the minimum reserved bandwidth and share the unreserved one. 1. Introduction Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 1] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 The Assured Forwarding PHB [RFC2597] group is intended to provide high probability packets forwarding to traffic not exceeding the subscribed profile (in-profile). Users are authorised to exceed the subscribed profile, however the excess traffic (out-of-profile) is not delivered with a high probability as the in-profile one. The data traffic is subject to marking in order to provide a low drop probability for in-profile traffic and a higher drop probability for out-of-profile one. The AF PHB specifies three drop precedences within each of four defined traffic classes. While the use of 3 levels of drop precedences was mainly proposed to prevent non responsive traffic from getting more than its fair share of the non reserved bandwidth, several contributions presented simulations where the use of only two drop precedences is enough to achieve the protection of TCP traffic [Goyal1, Goyal2]. In [Ibanez, Yeom] simulation results show that 2 drop precedences are not enough to fully utilize the reserved bandwidth. In this memo, we present some simulation configurations in which the use of three levels of drop precedences can be used to allow TCP traffic to achieve its reserved bandwidth. 2. Simulation configuration Our simulation configuration is depicted in Figure 2.1. The network is composed of 4 pairs of interconnected LANs. Each LAN is attached to a customer router (C1-C4, C1'-C4'). Customer routers are connected via a 139 Mbps link to the backbone composed of a single bottleneck link between the edge routers ER1 and ER2. In LAN1, LAN2 and LAN3 (connected respectively to C1, C2 and C3) a set of 10 TCP-SACK sources are performing infinite file transfer to destinations in LAN1', LAN2' and LAN3' (connected respectively to C1', C2' and C3'). The MSS for TCP is set to 536 Bytes. In LAN4 (connected to C4) a UDP source is sending fixed sized packets of 576 Bytes to a UDP destination in LAN4' (connected to C4'). The inter packets sending time is exponentially distributed. The mean UDP rate varies for the different simulations from 0 Mbps up to 60 Mbps. The customer routers perform traffic conditioning: packets are marked with the appropriate drop precedence according to a certain traffic profile. In order to compare the performance of two color and three color markers we use the Token Bucket Marker (TBM) allowing two drop precedences and the two rate Three Color Marker (trTCM) allowing three drop precedences [Heinanen]. With the TBM, tokens are generated at a constant rate, i.e. the Committed Information Rate (CIR), up to a maximum value, i.e the Committed Burst Size (CBS). When a packet arrives, the bucket occupancy is decremented by the packet size and the packet is marked with the drop precedence 0 if enough tokens are available. The packet is marked with the drop precedence 1 otherwise. In the latter case Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 2] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 the bucket occupancy is not decremented. With the trTCM, two buckets, with sizes CBS and Peak Burst Size (PBS), and two generation rates, CIR and Peak Information Rate (PIR), are used. Tokens in each bucket are generated at a constant rate: CIR and PIR respectively. Two modes of operation can be considered: the color blind and the color aware mode. In our simulation we used only the color blind mode in which each arriving packet is forwarded with the drop precedence 0 if the packet size is less than the first and the second bucket occupancy (both buckets are decremented by the packet size), with drop precedence 1 if the packet size is more than the the first bucket occupancy but less than the the second (the second bucket is decremented by the packet size) and finally with drop precedence 2 if the packet size is more than the first and the second bucket occupancy. 2.5 ms, 139 Mbps 2.5 ms, 139 Mbps <------------> <------------> \+---+ +---+/ -| C1|--------- ---------|C1'|- /+---+ | +----------+ +----------+ | +---+\ \+---+ |---| | | |---| +---+/ -| C2|------ | | | | ------|C2'|- /+---+ |------| | | |------| +---+\ \+---+ |------| ER1 |-----| ER2 |------| +---+/ -| C3|------ | | | | ------|C3'|- /+---+ |---| | | |---| +---+\ \+---+ | +----------+ +----------+ | +---+/ -| C4|--------- <-----> ---------|C4'|- /+---+ 10 ms, 34 Mbps +---+\ Figure 2.1: Simulation model The simulation parameters of the TBM and trTCM are, respectively, given in Table 2.1 and Table 2.2. We configured the simulation parameters for the trTCM such that all the excess traffic for TCP sources is marked with drop precedence 1 and all the excess traffic from the UDP source is marked with drop precedence 2. This is done by setting the PIR for customers running TCP sources to a very high value (equal to the link rate) and to the CIR for the customer running the UDP source. Table 2.1: Simulation parameters for TBM Scenario 1 Scenario 2 Scenario 3 CBS (C1, C2, C3, C4) 5 KB 10 KB 50 KB CIR (C1, C2, C3, C4) 2 Mbps 2 Mbps 2 Mbps Table 2.2: Simulation parameters for trTCM Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 3] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 Scenario 1 Scenario 2 Scenario 3 CBS (C1, C2, C3, C4) 5 KB 10 KB 50 KB PBS (C1, C2, C3, C4) 5 KB 10 KB 50 KB CIR (C1, C2, C3, C4) 2 Mbps 2 Mbps 2 Mbps PIR (C1, C2, C3) 139 Mbps 139 Mbps 139 Mbps PIR (C4) 2 Mbps 2 Mbps 2 Mbps Routers ER1 and ER2 implement the RIO algorithm [Clark] that we have extended to support three drop precedences instead of two. The parameter maxp was set to 0.02, 0.05 and 0.1 for drop precedence 0, 1 and 2 respectively. The thresholds (minth, maxth) in KB for drop precedences 0, 1 and 2 are respectively (45, 90), (90, 180) and (180, 360). For each drop precedence, i, an average queue length is calculated using the current buffer occupancy of all packets with drop precedences ranging from 0 to i. As in [Floyd], the weight to calculate the average queue length is set to 0.002. 3. Simulation Results In this section we present simulation results obtained with 2 and 3 drop precedences. Tables 3.1, 3.2 and 3.3 give the simulation results for Scenario 1, Scenario 2 and Scenario 3, respectively, described in Section 2. From Tables 3.1, 3.2 and 3.3 we can see that when only two drop precedences are used the TCP aggregates are, in most cases, only able to consume their reserved bandwidth (i.e. the CIR) and a portion of the unreserved bandwidth if the UDP sending rate is less than its CIR plus the unreserved bandwidth. For low values of CBS (Table 3.1) TCP aggregates are not able to fully utilize the CIR when the UDP sending rate exceeds its CIR plus the unreserved bandwidth. When 3 drop precedences are used the UDP source is only able to fully utilize its CIR. The unreserved bandwidth is fully utilized by TCP aggregates. In addition the obtained results do not vary when the CBS varies (Tables 3.1, 3.2 and 3.3). Table 3.1: Simulation results for Scenario 1. ------------------------------------------------------------------ | with two drop precedences | with three drop precedences UDP rate| Sum(TCP Throu.)| UDP Throu.| Sum(TCP Throu.)| UDP Throu. ------------------------------------------------------------------ 0 33.98 - 33.98 - 5 29.05 4.92 31.96 2.03 10 24.20 9.90 31.93 2.05 15 19.34 14.65 31.97 2.02 20 14.66 19.32 31.94 2.05 25 10.62 23.31 31.94 2.04 Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 4] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 30 7.80 26.11 31.93 2.06 35 6.34 27.56 31.94 2.04 40 5.60 28.30 31.93 2.06 45 5.29 28.61 31.93 2.05 50 5.22 28.70 31.93 2.04 55 5.14 28.77 31.92 2.06 60 5.01 28.91 31.93 2.05 Table 3.2: Simulation results for Scenario 2. ------------------------------------------------------------------ | with two drop precedences | with three drop precedences UDP rate| Sum(TCP Throu.)| UDP Throu.| Sum(TCP Throu.)| UDP Throu. ------------------------------------------------------------------ 0 33.98 - 33.98 - 5 29.06 4.91 31.95 2.03 10 24.18 9.79 31.95 2.03 15 19.29 14.68 31.95 2.03 20 14.71 19.32 31.93 2.05 25 10.60 23.38 31.95 2.03 30 7.96 25.94 31.90 2.07 35 6.81 27.08 31.94 2.04 40 6.36 27.53 31.94 2.04 45 6.15 27.74 31.95 2.02 50 5.97 27.92 31.91 2.06 55 5.87 28.01 31.90 2.08 60 5.78 28.50 31.93 2.05 Table 3.3: Simulation results for Scenario 3. ------------------------------------------------------------------ | with two drop precedences | with three drop precedences UDP rate| Sum(TCP Throu.)| UDP Throu.| Sum(TCP Throu.)| UDP Throu. ------------------------------------------------------------------ 0 33.98 - 33.98 - 5 29.03 4.97 31.97 2.04 10 24.15 9.85 31.94 2.05 15 19.33 14.67 31.97 2.04 20 14.69 19.31 31.95 2.06 25 10.52 23.43 31.97 2.04 30 7.89 26.05 31.94 2.07 35 6.90 27.02 31.95 2.06 40 6.52 27.40 31.93 2.05 45 6.29 27.61 31.93 2.07 50 6.25 27.66 31.91 2.09 55 6.19 27.71 31.92 2.09 60 6.08 27.80 31.91 2.09 4. Simulations with large RTT and large CIR Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 5] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 In this paragraph we study the effect of having a large RTT (large propagation delay) and a large value for the CIR. The simulation model studied is similar to the one presented in paragraph 2. The only diffrences are: the propagation delay of the link between customer 2 and ER1 is set to 100 ms instead of 2.5 ms and customer 3 has a CIR of 20 Mbps instead of 2 Mbps. The UDP source is sending fixed sized packets of 576 Bytes with an inter packet sending time exponentially distributed. The mean rate is set to 25 Mbps for all simulations. We performed 5 sets of simulations with the following CBS values: 5, 10, 50, 100, 500 KB. The simulation parameters are summarized in Table 4.1. The remaining parameters are set as in section 2. We also performed simulations with the Rate Adaptive Shaper (RAS) [Bonaventure] which is combined with the TBM and the trTCM. The RAS is used in front of the marker and aims at increasing the number of low drop precedence packets by reducing the burstiness of the traffic. Tables 4.2, 4.3, 4.4 and 4.5 give the simulation results. From Table 4.2 we can see that with only two drop precedences customer 3 is never (for the used values of CBS) able to use its reserved bandwidth (i.e. CIR). Customer 2 succeeds to use its CIR for high values of the CBS (50, 100, 500 KB). When 2 drop precedences are used in combination with the RAS (see Table 4.3) mostly all customers, except customer 3 when the CBS is too small, can fully utilize their CIR. The use of the shaper allows the customers to obtain an aggregate throughput close to the CIR but the unreserved bandwidth is mostly utilized by the UDP source. When 3 drop precedences are used (see Table 4.4) it appears clearly that all customers obtain an aggregate throughput which is higher than the CIR. The unreserved bandwidth is mostly used by the TCP aggregates. In addition the obtained aggregate throughput do not vary a lot when the CBS vary. Finally the use of three drop precedences combined with the RAS gives results very close to ones obtained without the RAS for this configuration. From this set of simulations we have shown the following properties: - The use of 2 drop precedences do not allow, in the presence of UDP traffic, the TCP aggregates to utilize the unreserved bandwidth which is mostly consumed by the UDP source. TCP aggregates can fully utilize their CIR when the appropriate CBS values are used. The use of the RAS can help TCP aggregates to utilize their CIR much better but not the unreserved bandwidth. Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 6] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 - The use of 3 drop precedences allows the TCP aggregates to share the unreserved bandwidth. In addition the obtained throughput does not vary a lot with CBS variations. Table 4.1: Simulation parameters customer 1: CIR of 2 Mbps customer 2: CIR of 2 Mbps, large RTT customer 3: CIR of 20 Mbps customer 4: CIR of 2 Mbps, UDP of 25 Mbps CBS : 5, 10, 50, 100, 500 KB Table 4.2: Simulation results with 2 drop precedences CBS customer 1 customer 2 customer 3 customer 4 5 2.39 1.65 13.59 16.35 10 2.36 1.86 14.05 15.71 50 2.21 2.00 18.42 11.35 100 2.21 2.02 18.53 11.22 500 2.26 2.12 19.41 10.20 Table 4.3: Simulation results with 2 drop precedences - shaper CBS customer 1 customer 2 customer 3 customer 4 5 2.09 1.99 19.07 10.84 10 2.08 1.99 19.32 10.59 50 2.05 1.99 19.83 10.11 100 2.10 1.99 19.80 10.08 500 2.05 1.99 19.83 10.11 Table 4.4: Simulation results with 3 drop precedences CBS customer 1 customer 2 customer 3 customer 4 5 6.80 2.98 21.96 2.23 10 6.65 2.96 22.22 2.15 50 6.54 3.03 22.14 2.28 100 6.36 3.05 22.23 2.34 500 6.37 2.97 22.46 2.19 Table 4.5: Simulation results with 3 drop precedences - shaper CBS customer 1 customer 2 customer 3 customer 4 5 6.64 3.36 21.72 2.26 10 6.01 3.56 22.19 2.23 50 6.03 3.54 22.15 2.26 100 6.23 3.40 22.12 2.23 500 5.89 3.64 22.18 2.27 5. Conclusion Simulation results presented in this memo can be summarized as follows: Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 7] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 - When only two drop precedences are used TCP aggregates can, in most cases, fully use their reserved bandwidth if the markers' parameters are appropriately set (CBS value). However the unreserved bandwidth is mostly used by UDP sources. - The use of three drop precedences allows the TCP aggregates to fully use their reserved bandwidth and to share the unreserved one. In addition the obtained throughputs do not vary a lot with the variations of the CBS parameter. However the unreserved bandwidth is not fairly distributed among the TCP aggregates. 6. References [Bonaventure] O.Bonaventure, S. De Cnodder, A rate adaptive shaper for dif- ferentiated services. Internet draft, IETF, draft-bonaventure- diffserv-rashaper-00.txt. Work in progress. [Clark] D. D. Clark, W. Fang, Explicit Allocation of Best-Effort Packet Delivery Service. IEEE/ACM Trans. on Networking, Vol. 6, No. 4, August 1998. [Floyd] S. Floyd, V. Jacobson, Random Early Detection Gateways for Congestion Avoidance. IEEE/ACM Transactions on Networking, August 1993. [Ibanez]J. Ibanez, K. Nichols, Preliminary Simulation Evaluation of an Assured Service. Internet draft, IETF, draft-ibanez-diffserv- assured-eval-00.txt. Work in progress. [Goyal1]M. Goyal, P. Misra, R. Jain, Effect of Number of Drop Pre- cedences in Assured Forwarding. Internet draft, IETF, draft- goyal-dpsty-diffserv-00.txt. Work in progress. [Goyal2]M. Goyal, P. Misra, R. Jain, Effect of number of drop pre- cedences in Assured Forwarding. Available from http://www.cis.ohio-state.edu/~ jain/papers/dpstdy_globecom99.htm [Heinanen] J. Heinanen, and R. Guerin, A Two Rate Three Color Marker. Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 8] Internet Draft draft-elloumi-diffserv-threevstwo-00.txt June 1999 Internet draft, IETF, draft-heinanen-diffserv-trtcm-01.txt. Work in progress. [RFC2597]J. Heinanen, F. Baker, W. Weiss, J. Wrockawski, Assured For- warding PHB Group. RFC 2597, June 1999. [Yeom] I. Yeom, A. L. N. Reddy, Realizing throughput guarantees in a differentiated services network. IEEE ICMCS'99, June 1999. Acknowledgments The authors would like to thank Jordi Nelissen for several dis- cussions. Authors' Addresses Omar Elloumi Alcatel Corporate Research Center Fr. Wellesplein 1, B-2018 Antwerpen, Belgium. Phone : 32-3-240-7833 Fax : 32-3-240-9932 E-mail: Omar.Elloumi@alcatel.be Stefaan De Cnodder Alcatel Corporate Research Center Fr. Wellesplein 1, B-2018 Antwerpen, Belgium. Phone : 32-3-240-8515 Fax : 32-3-240-9932 E-mail: stefaan.de_cnodder@alcatel.be Kenny Pauwels Alcatel Corporate Research Center Fr. Wellesplein 1, B-2018 Antwerpen, Belgium. Phone : 32-3-241-5941 Fax : 32-3-240-9932 E-mail: kenny.pauwels@alcatel.be Elloumi, De Cnodder, Pauwels Drop precedences in AF [Page 9]