Transport Meeting Notes 2015 Oct 15


1. Naming of congestion window sysctl

Suggest taking Robert's suggestion. Hiren will wait 48 hours for other comments and then commit.

2. mbuf safety net

Jonathan has proposed a change that we would run out of mbufs This only runs when you hit the max on a UMA zone.

This is separate from a tcp_drain call as it addresses the entire system.

Benno says that they are looking at having pools to buffer this.

3. Reviewers

Hiren has posted several reviews that need to be looked at.

George to create a transport group on the phabricator/review site.

4. TCP Tracing from Sewell et al.

They are now using 11-CURRENT and want to try the DTrace tracepoints that have been added to TCP on FreeBSD.

Their long term strategy is to update their model to understand our more modern network stack.

5. New Business

Try to get Swinburne folks to talk to us about MPTCP. MPTCP updated but not yet reviewed. They have a test bed. Hiren to direct them to the transport@ list.

DCTCP is in HEAD and there are improvements coming.

Will gnn remove the TCP_DEBUG from the kernel now that we have DTrace probes? rwatson leans in favor keeping it and gnn agrees.

Jonathan proposed turning TCP into a module by using function pointers per socket. Randall points out that they're already doing this in a limited manner. The idea is to have per socket TCP stacks via a library of TCP functions. Others are generally interested in this. Jonathan will start a sub page on the Wiki so that we can design this in the clear. Randall will clean up his patch and put it up for review, as it is a limited move in the new direction.

Can we flag things are experimental in Phabricator? There is a flag, but it's not, perhaps, strong enough.

15Oct2015 (last edited 2015-10-15 16:55:49 by GeorgeNevilleNeil)