Possibly issue with /32 vs. /24 wrt route caching from (net@) - UPDATE (hiren 04/21/2016 13:04 PDT) It turned out to be something broken with lacp/lagg and not route caching.
UDPv6 might also need the route caching changes as well. UDPv4 seems to be holding the lock longer now as well.
LLNW is looking at building out a client device lab. A hot issue has to do with under performance with a specific test where streams have .1% packet packet loss. Start with .1% and increment by .1%. Time to first re-buffer is the measurement. They will share results once they have them.
The patch from jtl@ does not solve all the underflow problems but does solve some of them. Hiren is looking at tcp_send_probe() at the end of tcp_output(). Hiren will continued to dig into this.
Any thoughts on Hiren's cwnd patch? Has not tried it in production but Randall said that dropping to 1 was better for his test than dropping to 1/2.
Hiren has a patch that enables the pipe calculation and cwnd via a sysctl. He will put that into Phabricator and then eventually in HEAD. This will make it interact correctly with RFC 6675.
Kevin will start getting data collection at their edge. He believes that they will have that soon.
George has started working with a group on a set of protocol conformance tests which ought to be open sourced in the near future and which cover the layers from NICs up to TCP.
Michael has an updated version of the packetdrill port which he will point us at so that George can update the port.