09 Feb 2017 Hangout Notes
- rrs@ mentions that they now have a "black box recorder" which is done by jtl@. This uses a set probes to collect data over time. It collects data in pcap form t so that you can look at it tcpdump. It can either stream data or dump data following a signal. jtl@ will put up a review soonish. gnn@ wonders if much can be done with DTrace.
- BBR has been handed over to the production team and will be off by default in the NF kernels, but can be turned on via a socket option. The performance is looking good in their testing. rrs@ wants to get this into their transport lab for testing. One issue remains the CPU utilization of the callout system. rrs@ will move BBR to using the pacing infrastructure as it has a lower CPU overhead than the callout system.
People need to review https://reviews.freebsd.org/D9356 which glebius has put up for review. This splits listen sockets from data transport sockets.
- rrs@ points out that if you are always in sequence and not having loss you will grow your window but will never tell the peer. He says he will make a patch for review.
- hiren@ is working with lstewart@ on what's broken with the fast recovery fix. The fix from a few months ago to not revert to 1MSS seems to result in a higher level of retransmits.
- gnn@ asks if tuexen@ is ready to push his changes to tcp-testsuite
- gnn@ will kick off a conversation with Neal at Google and tuexen@ about upstreaming to the main repo
rstone@ asks for review of https://reviews.freebsd.org/D9465 . bz@ suggests taking some of the UTO spares.
- tuexen@ asks what happens in IPv6 when we process a "packet too big" ICMP error. On IPv4 the host cache is updated. bz@ will send the things he knows about this directly to tuexen@
- rstone@ points out that clearing the SRTT after 3 retransmits makes debugging harder. rstone@ will put up phabricator on this topic.
tuexen@ asks about this review https://reviews.freebsd.org/D9475 which addresses an issue with very small receive buffers.