Bandwidth for High Speed Networks

As we increase the latency of the congested link redi

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: on algorithm described in 8 , which we call RED with Identi cation REDI.13 In each case there are multiple ows traversing a 1.5 Mbps link with a latency of 3 ms; the output bu er size is 32 KB and all constants K , K , and Kc, respectively, are set to 400 ms. Table 5 shows the bandwidth allocations under REDI and CSFQ averaged over 100 sec. In the rst experiment Simulation 1, we consider a 1 Mbps UDP ow and two TCP ows; in the second experiment Simulation 2 we have a standard TCP TCP-1 and a modi ed TCP TCP-2 that opens the congestion window three times faster. In both cases REDI fails to identify the unfriendly ow, allowing it to obtain almost two-thirds of the bandwidth. As we increase the latency of the congested link, REDI starts to identify unfriendly ows. However, for some values as high as 18 ms, it still fails to identify such ows. Thus, the identi cation approach still awaits a viable realization and, as of now, the allocation approach is the only demonstrated method to deal with the problem of unfriendly ows. 4.3 Punishment Earlier in this section we argued that the allocation approach gave drop-intolerant ows an incentive to adopt endto-end congestion control. What about drop-tolerant ows? We consider, for illustration, re-hose applications that have complete drop-tolerance: they send at some high rate and get as much value out of the fraction of arriving packets, call it x, as if they originally just sent a stream of rate x . That is, these re-hose applications care only about the ultimate throughput rate, not the dropping rate.14 In a completely static world where bandwidth shares were constant such re-hose" protocols would not provide any advantage over just sending at the fair share rate. However, if the fair shares along the path were uctuating signi cantly, then re-hose protocols might better utilize instantaneous uctuations in the available bandwidth. Moreover, re-hose protocols relieve applications of the burden of trying to adapt to their fair share. Thus, even when restrained to their fair share there is some incentive f...
View Full Document

Ask a homework question - tutors are online