There always seems to be a steady stream of proposals for changes to TCP congestion control. Some are more reasonable than others of course, but it seems to me that its pretty common for the same sorts of issues to crop up time and again. So I’ve put together a page which tries to document a few issues that have already been found (sometimes after painful experience) to be problematic and included references as entry points into the literature.
Archive for the ‘TCP’ Category
Some known ways to break TCP
Friday, May 15th, 2009Swinburne TCP Tests over DSL links
Thursday, May 14th, 2009The folks at Swinburne reported this week on tests comparing the performance of TCP variants over home DSL lines.
In summary, they consider a home DSL link and compare the impact of Reno, Cubic and HTCP on competing voice or games traffic. The find that Cubic is the most aggressive in its impact on the low rate traffic due to a combination of:
(a) Its concave cwnd shape
(b) Backing cwnd off by only 0.8 on loss (rather than by 0.5 in Reno, and also by HTCP in this sort of situation)
(c) Its behaviour at lower cwnds (where HTCP reverts directly to Reno while cubic takes a different approach).
This work appears in the paper Collateral Damage: The Impact of Optimised TCP Variants On Real-time Traffic Latency in Consumer Broadband Environments presented at Networking 2009.
We also carried out some tests at the Hamilton Institute a while back on flow completion times over DSL links and found that the slow convergence of Cubic can penalise later starting flows.