Weekly Hangout Notes 23 Feb 2017
- hiren@ is looking into an issue where not going back to slow start causes a lot retransmits. If this issues cannot be addressed without further scaffolding hiren@ will revert the change made a few months back. In part the issue is due to the stack not consider packets in flight. The work that is needed to support this correctly is already in the pipeline to be committed.
- Using the tcp-testsuite gnn@ found that changing the maxsockbuf up and back makes it so that you can no longer create sockets. rstone@ or gnn@ may address it. This only effects systems where the socket buffers are set to be very large by default.
- rstone@ would like to get his low latency code into the main stack. The code needs a cleanup and then rstone@ would like comments on the code. The changes mostly relate to low RTO and delayed acks. The TCP timestamp needs to have higher precision than ticks on low RTT connections. For the low latency sockets the connecting socket will change its clock source and for listen sockets the accepted socket will inherit the better clock. jtl@ points out that there is a draft at the IETF on changing this timestamp as well.
jtl@ asks folks to review https://reviews.freebsd.org/D9485
- jtl@ notes that the hashing in lagg currently is not well balanced. They are thinking of changing the hashing algorithm to look at more bits and hash across queues instead of ports. Two people who were recommended to jtl@ are adrian@ and npn@. teuxen@ will forward a paper on this topic.
glebius@ would like a review of https://reviews.freebsd.org/D9770 which is a work in progress and not for a general commit to HEAD. He asks that teuxen@ give some help with the SCTP parts of this particular patch. teuxen@ and glebius@ will discuss by email
- jtl@ is looking into the problem of not canceling interrupts when canceling a callout. rstone@ has also looked into this. It's important to the stack but the underlying code is difficult to work with.