Figure 6:(a) The normalized throughput of UDP-0 as a function of the number of congested links. (b) The same plot when UDP-0 is replaced by a TCP flow FRED RED 53221436 5452 FIFO Table 1: Statistics for an ON-OFF flow with 19 competing Table 3: The mean throughputs (in packets) and st andard TCPs Alows(all numbers are in packets) deviations for 19 TCPs in the presence of a UDP How along a link with propagation delay of 100 ms, The UDP sends at the link capacity of 10 Mbps. gorithm mean time std dev is 64 KB, we set constants K, Ka, and Ke to be 250 ms i. e, about two times larger than the maximum queue de- I RED 5921274 lay. An interesting point to notice is that, unlike DRR and FIFO 8401695 CSFQ, Fred does not provide fair bandwidth allocation in this scenario. Again, as discussed in Section 3. 2, this is Table 2: The mcan transfcr times (in ma)and the corre due to the fact that RLM and TCP use different, end-to-end sponding standard deviations for 60 short TCPs in the pres- congestion control algorith ce of a UDP flow that sends at the link capacit Mbp 3. 4 Different Traffic Models So far we have only considered UDP, TCP and layered mi ticast, traffic sources. We now look at two additional source 3.3 Coexistence of Different Adaptation Schemes models with greater degrees of burstiness. We again con- In this experiment we investigate the extent to which CSFQ gle 10 Mbps congested link. In the first exper iment, this link is shared by one ON-OFF source and 19 Receiver-driven Layered Multicast(RLM)(15] is an adaptive TCPs.. The ON and OFF periods of the ON-OFF source into a number of layers (each to its own multicast group)and of 100 ms and 1900 ms respectively. During the ON perla the receiver joins or leaves the groups associated with th the ON-OFF sourcc sends at 10 Mbps. Notc that thc ON layers based on how many packet drops it is experiencing the der as the averaging intervals K e consider a 4 Mbps link traversed by one TCP and three and Kc which are all 100 ms, so this experiment is designed RLM flows. Each source uses a seven layer encoding, where to test to what extent CSFQ act over short timescales RLM case this will correspond to each receiver subs.E layer i sends 2+4 Kbps; each layer is modeled by a UDP The ON-OFF source sent 6758 packets over the course of affic source. The fair share of each How is 1Mbps the experiment. Table 1 shows the number of packets from the ON-OFF source dropped at the congested link. The to the first five layers? DRR results show what happens when the ON-OFF source The receiving rates averaged over I second interval for restricted to its fair share at all times. FRED and CSFQ each algorithm are plotted in Figure 7. Since in this experi- also are able to of fairnes ment the link bandwidth is 4 Mbps and the router buffer size ulates Web traffi 60 TCP transfers -arrival times are More precisely, we have >s,2+4 Kbps = 0.992 Mbps distributed with the mean of 0.05 ms, and theFigure 6: (a) The normalized throughput of TJDP-0 as a function of the number of congested links. (b) The same plot when UDP-0 is replaced by a TCP flow. Algorithm delivered dropped DRR I 601 1 6157 t CSFQ: I 1680 I 5078 1 Table 1: Statistics for an ON-OFF flow with 19 competing TCPs flows (all numbers are in packets). FIFO 840 1 1695 Table 2: The mean transfer times (in ms) and the corre￾sponding standard deviations for 60 short TCPs in the pres￾ence of a UDP flow that sends at the link capacity, i.e., 10 Mbps. 3.3 Coexistence of Different Adaptation Schemes In this experiment we investigate the extent to which CSFQ can deal with flows that employ different adaptation schemes. Receiver-driven Layered Multicast (RLM) [15] is an adaptive scheme in which the source sends the information encoded into a number of layers (each to its own multicast group) and the receiver joins or leaves the groups associated with the layers based on how many packet drops it is experiencing We consider a 4 Mbps link traversed by one TCP and three RLM flows. Each source uses a seven layer encoding, where layer i sends 21t4 Kbps; each layer is modeled by a UDP traffic source. The fair share of each flow is 1Mbps. In the RLM case this will correspond to each receiver subscribing to the first five layers’. The receiving rates averaged over 1 second interval for each algorithm are plotted in Figure 7. Since in this experi￾ment the link bandwidth is 4 Mbps and the router buffer size ‘More precisely, we have c:=, 2’t4 Kbps = 0.992 Mbps. Algorithm 1 mean std. dev DRR I 6080 I 64 CSFQ 5761 220 FRED 4974 190 RED 628 80 FIFO 378 69 Table 3: The mean throughputs (in packets) and standard deviations for 19 TCPs in the presence of a UDP flow along a link with propagation delay of 100 ms. The UDP sends at the link capacity of 10 Mbps. is 64 KB, we set constants I(, li,, and h’, to be 250 ms, i.e., about two times larger than the maximum queue de￾lay. An interesting point to notice is that, unlike DRR and CSFQ, FRED does not provide fair bandwidth allocation in this scenario. Again, as discussed in Section 3.2, this is due to the fact that RLM and TCP use different end-to-end congestion control algorithms. 3.4 Different Traffic Models So far we have only considered UDP, TCP and layered mul￾ticast traffic sources. We now look at two additional source models with greater degrees of burstiness. We again con￾sider a single 10 Mbps congested link. In the first exper￾iment, this link is shared by one ON-OFF source and 19 TCPs. The ON and OFF periods of the ON-OFF source are both drawn from exponential distributions with means of 100 ms and 1900 ms respectively. During the ON period the ON-OFF source sends at 10 Mbps. Note that the ON￾time is on the same order as the averaging intervals K, K,, and I(, which are all 100 ms, so this experiment is designed to test to what extent CSFQ can react over short timescales. The ON-OFF source sent 6758 packets over the course of the experiment. Table 1 shows the number of packets from the ON-OFF source dropped at the congested link. The DRR results show what happens when the ON-OFF source is restricted to its fair share at all times. FRED and CSFQ also are able to achieve a high degree of fairness. Our next experiment simulates Web traffic. There are 60 TCP transfers whose inter-arrival times are exponentially distributed with the mean of 0.05 ms, and the length of each 125