Wednesday, September 10, 2008

A third look at congestion avoidance

This time we've moved from the edge into the gateways.

Their idea was great: let's run simulations to show that our modified version of Fair Queueing can outperform traditional congestion control mechanisms.

Much of the writing was comprehensible and well justified. It seems to me that this work was low hanging fruit. Though their categorization of the flaws in Nagle's version of Fair Queuing and the solutions they provide, supported by simulations which show large performance gains, were certainly well done and non-trivial.

I feel that they used math which was overly complex and under-explained in the performance section. For the majority of the paper, I was unsure what they meant by fairness with regards to promptness allocation, and a section of dense math didn't help much. I had to wait till the end of the section to get an understandable summary of the problem they were solving (i.e. "This ability to provide lower delay to lower throughput sources completely independent of the window sizes of the FTPs is one of the most important features of fair queuing") Also, Figure 1 was poorly explained.

1 comment:

Randy H. Katz said...

I have to agree about your comment on the density of the mathematics -- and the introduction of the mystery parameter delta which allows flows that underutilize their allocation to get priority dispatch. Further, as the second paper indicates, FQ was infrequently implemented in routers, although now it is more available (it is unclear however how much it used).