Transport Meeting Notes 11 Feb 2016
Robert, Bjoern, Hannes, Michael, Kevin, Hiren, Ryan, Jason W, Jason E. George
Robert asks about SCTP and TCP congestion control algorithms and whether they are shared. MPTCP would be easier but SCTP probably not so much.
Hiren has put his congestion window change on hold pending Lawrence.
Hiren will talk to Randall about projects branch for 10 for alternate stacks and other HEAD based work.
Hiren will check on whether or not we set PSH on close. This relates to the RTO problem from the last meeting. LRO may interfere with this.
Discussion of how we deal with LRO generally. Right now LRO somewhat interferes with ECN according to Ryan.
Hannes has been working with the FreeBSD stack formal model. He has found the traces from 10 years ago as well as the code. He can reproduce the results on the old traces. He plans to restart this work on a more recent (HEAD?) starting in 2 weeks. He will share the traces on the mailing list. The github code is here: https://github.com/PeterSewell/netsem The traces are here http://www.cl.cam.ac.uk/~hm519/netsem/
Kevin has asked about the XNU code for Fast Open. Robert points out that the code is not under the friendliest license. Kevin will send a link to the Fast Open code in the XNU tree.
Hiren asks getting 4898 data out and why there is no standard. For FreeBSD, George says that it makes sense to pull out via sysctl. George will work with Hiren to expose this all via DTrace.
There is a patch for ifLib that MattM has put up for review on Phabricator. Andrew Gallatin is actively reviewing it and talking to Intel as well.
In low latency networks we want to scaled down the RTO to deal with low RTT. The assumption is that we can only control the sender. Matt will look at the PUSH flag implementation as well. What do you do about TSO? If on the last packet Matt is tracking packets sent to trigger the ACK correctly.
The route caching review is stuck but Melifaro may be able to break the logjam.