Unformatted text preview: xt of the identi cation approach is far from a solved problem; the challenge
is to determine if a ow has decreased usage in response to
increases in overall packet drop rates 8 .
Identifying unresponsive ows is more straightforward
in the allocation approach, since here one need only determine if a ow has had signi cantly high drop rates over a
long period of time. As a proof of concept we have implemented a simple identi cation and punishment mechanism.
First, we examine o -line the last n dropped packets and
then monitor the ows with the most dropped packets. Second, we estimate the rate of each of these monitored ows;
when a ow's rate is larger than a a 1, we start
dropping all of its packets. Third, we continue to monitor
penalized ows, continuing punishment until their arrival
rate decreases below b b 1. Using the parameters
a = 1:2, b = 0:6, and n = 100, we applied this algorithm to
Simulation 1 in Table 5; the UDP ow was identi ed and
penalized in less than 3 seconds. Our task was easy because
the identi cation of unresponsive ows can be based on the
result packet drops over long periods of time rather than
on trying to examine the algorithm detecting whether it
actually decreased its rate in response to an increase in the
drop rate. Note also that the allocation approach need only
distinguish between responsive and unresponsive in the punishment phase, an inherently easier task than distinguishing
friendly from unfriendly.
In summary, to provide incentives for drop-tolerant ows
to use responsive end-to-end congestion control, it may be
necessary to identify, and then punish, unresponsive ows.
overhead associated with them, and it isn't clear whether, in practice,
the overheads are greater or less than the advantages gained. However, one can certainly not claim, as we did above for drop-intolerant
applications, that the allocation approach gives drop-tolerant applications a strong incentive to use responsive end-to-end congestion
16 Another possible method, used in ATM ABR, is...
View Full Document
- Fall '10
- Eugene Ng
- Scheduling algorithm, Round-robin scheduling, Scheduling algorithms, Ow, Fair queuing, ows