Transport Group Meeting 24 Sep 2015
Problems with loss have been confirmed by Hire and Randall.
Hiren to add bugzilla bugs for the issues found thus far.
Discussions are now moving to the new transport@ mailing list.
Silby's tcp tool is now on github (see Tools section of main page) This is a too for finding TCP regressions that was presented at BSDCan 2009. There are questions as to whether the tool will gracefully handle updates to the TCP stack or if changes will always cause false positives. Packet traces can be archived to protect against a loss of history.
We then went on to discuss Peter Sewell et al's TCP formalization work from 10 years ago. The code is also up on github but under an license that is difficult to work with. Peter and his group are working through the license issues.
gnn has now started an RFC list (see main page).
The DTrace TCPDEBUG tracepoints are now in HEAD.
gnn to write a wiki page/document on "how to submit a TCP bug"
Various tools were discussed, including some dummynet work by Lawrence Stewart which is now out of date, Hiren's own dummynet hacks, as well as tcpreplay. The NS3 network simulator was mentioned as well.
A major topic of concern is that of how the stack handles Out of Order packets. Robert mentioned that Mark Handley at UCL and Grenville Armitage at Swinburne should both be contacted on this topic. Many Out of Order issues are due to driver bugs, notably in the Intel NIC drivers. George is going to setup a set of DTrace SDTs and a process for adding them to drivers so that these problems are easier to detect. George will work with Sean Bruno to get these SDTs into the Intel drivers. (Now in process).
Another problem is that our congestion control code often fights with the rest of the TCP stack. Bugs need to be opened on this topic.
Hiren mentioned Laminar, from Google, which is something that the group ought to look at but which will require a serious undertaking to get going.