{[ promptMessage ]}

Bookmark it

{[ promptMessage ]}

Bandwidth for High Speed Networks

In general we nd that csfq achieves a reasonable

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: esenting the unfair status quo. The goal of these experiments is to determine where CSFQ sits between these two extremes. FRED is a more ambiguous benchmark, being somewhat more complex than CSFQ but not as complex as DRR. In general, we nd that CSFQ achieves a reasonable degree of fairness, signi cantly closer to DRR than to FIFO or RED. CSFQ's performance is typically comparable to FRED's, although there are several situations where CSFQ signi cantly outperforms FRED. There are a large number of experiments and each experiment involves rather complex dynamics. Due to space limitations, in the sections that follow we will merely highlight a few important points and omit detailed explanations of the dynamics. 3.1 A Single Congested Link We rst consider a single 10 Mbps congested link shared by N ows. The propagation delay along the link is 1 ms. We performed three related experiments. In the rst experiment, we have 32 UDP ows, indexed from 0, where ow i sends i + 1 times more than its fair share of 0.3125 Mbps. Thus ow 0 sends 0.3125 Mbps, ow 1 sends 0.625 Mbps, and so on. Figure 3a shows the average throughput of each ow over a 10 sec interval; FIFO, RED, and FRED-1 fail to ensure fairness, with each ow getting a share proportional to its incoming rate, while DRR is extremely e ective in achieving a fair bandwidth distribution. CSFQ and FRED-2 achieve a less precise degree of fairness; for CSFQ the throughputs of all ows are between ,11 and +5 of the ideal value. In the second experiment we consider the impact of an ill-behaved UDP ow on a set of TCP ows. More precisely, the tra c of ow 0 comes from a UDP source that sends at 10 Mbps, while all the other ows from 1 to 31 are TCPs. Figure 3b shows the throughput of each ow averaged over a 10 sec interval. The only two algorithms that can most e ectively contain the UDP ow are DRR and CSFQ. Under FRED the UDP ow gets almost 1.8 Mbps close to six times more than its fair share while the UDP only gets 0.396 Mbps and 0.361 Mbps under DRR and CSFQ, respectively. As expected FIFO and RED perform p...
View Full Document

{[ snackBarMessage ]}