Meeting Notes 5 May 2016

Issues with the route cache addition and the flowtable. The panic has been posted to net@

The 6675 code will not be abandoned but will not be committed as yet because it is only a partial implementation.

Intel remains committed to having ifLib support in 11. Can we bring in ifLib without a driver? Drew will commit this to HEAD soon.

Hiren is working with various congestion control algorithms for a particular workload. CDG is not used that much so has issues which Hiren is addressing. He is keeping records of the experiments that we can look at later. He also has a patch for the ssthresh interaction that he'll put into Phabricator and share with others.

Ryan is working on incast performance. Will put it into both their internal code base as well as upstreaming it to FreeBSD. Fixing this likely requires improving the timer system as well. Can't just turn on everywhere because of delayed ack interactions. This may need a new TCP and Socket option. Michael points out that there is already a similar option in SCTP.

In TCP performance tests Ryan sees mbuf allocation/deallocation as up to 20% of the overhead. Drew also has confirmed this with VTune as well. Rearranging some fields helped Drew's tests. It may be that rearranging the structure will improve cache performance.

Issues with IPv6 Route Advertisements are causing problems for some folks. There is a race condition that the Limelight setup exposes. This is a locking issue that leads to a panic(). George suggests a PCS script to generate RAs and Jonathan says he has something like that as well.