Unformatted text preview: oorly, with
the UDP ow getting over 8 Mbps in both cases.
In the nal experiment, we measure how well the algorithms can protect a single TCP ow against multiple Figure 5: Topology for analyzing the e ects of multiple congested links on the throughput of a ow. Each link has
ten cross ows all UDPs. All links have 10 Mbps capacities. The sending rates of all UDPs, excepting UDP-0, are
2 Mbps, which leads to all links between routers being congested.
ill-behaved ows. We perform 31 simulations, each for a
di erent value of N , N = 2 32. In each simulation we
take one TCP ow and N , 1 UDP ows; each UDP sends
at twice its fair share rate of 10 Mbps. Figure 4 plots the
ratio between the average throughput of the TCP ow over
10 seconds and the fair share bandwidth it should receive
as a function of the total number of ows in the system N .
There are three points of interest. First, DRR performs very
well when there are less than 22 ows, but its performances
decreases afterwards. This is because the TCP ow's bu er
share is less than three bu ers, which signi cantly a ects
its throughput. Second, CSFQ performs better than DRR
when the number of ows is large. This is because CSFQ is
able to cope better with the TCP burstiness by allowing the
TCP ow to have several packets bu ered for short time
intervals. Finally, across the entire range, CSFQ provides
similar or better performance as compared to FRED. 3.2 Multiple Congested Links We now analyze how the throughput of a well-behaved ow
is a ected when the ow traverses more than one congested
link. We performed two experiments based on the topology
shown in Figure 5. All UDPs, except UDP-0, send at 2
Mbps. Since each link in the system has 10 Mbps capacity,
this will result in all links between routers being congested.
In the rst experiment, we have a UDP ow denoted
UDP-0 sending at its fair share rate of 0.909 Mbps. Figure 6a shows the fraction of UDP-0's tra c that is forwarded versus the number of congested links. CSFQ and
FRED perform reasonabl...
View Full Document
- Fall '10
- Eugene Ng
- Scheduling algorithm, Round-robin scheduling, Scheduling algorithms, Ow, Fair queuing, ows