Attendees
- glebius@
- olivier@
- Peter Lei
- rscheff@
- thj@
- tuexen@
Ongoing Short Term Projects
glebius@ asks whether it makes sense to improve the TCP timer precision. tuexen@ states that this is very useful for testing purposes. glebius@ will provide a patch to tuexen@ for testing and tuexen@ will provide feedback, if such a patch us useful for FreeBSD in general. This applies to the default stack only, not the the RACK or BBR stack, which uses HPTS for timers. Setting the sysctl variable kern.timecounter.alloweddeviation to zero allows the timers to run of very precisely, which is useful when running packetdrill. However, it affects the overall system performance.
- At least for testing purposes (packetdrill) it would be good to have LRO support for tun interfaces. tuexen@ suggested to generalize LRO to not depend on an Ethernet header being present. gallatin@ prefers to not make the LRO code more complex, since it most likely will impact the performance. tuexen@ has implemented LRO support for the tap interfaces. packetdrill support will be committed soon.
- BBLog is missing a man page. tuexen@ will write one.
- tuexen@ will add tests for handling RST ACK segments in all states and test the base, RACK and BBR stack.
rscheff@ wants feedback for D41715, which is based on a patch from VMWare for CUBIC. cc@ will test and review and rscheff@ will integrate the feedback. Also cc@ should look at D41672.
Not clearing the score board after the first timer based retransmission, results in uncovering several old bugs on SACK enabled connections. rscheff@ wants review of D43355 and D43470. tuexen@ will review it.
PR 146845. glebius@ will look at it.
tuexen@ is addressing issues in the TCP stack reported by Coverity. tuexen@ wants feedback on D44584 and D44616.
Ongoing Longer Term Projects
- Now the RACK and the BBR stack are compiled by default in head. rrs@ will remove the BBR stack, since it has several limitations. tuexen@ will mark it as deprecated. We discussed if the RACK stack could be made the default and what should drive this decision. tuexen@ will reach out to the mailing list to ask people to test the RACK stack and provide feedback. In January, we will look again at the decision process based on the feedback we got. Feedback received:
- You need to compile in the TCPHPTS. It is a loadable module now, thanks to glebius@.
- RACK is on the loopback interface slower than the base stack (CPU limited). olivier@ has a script to test combinations of stacks and CC modules.
A problem with RACK an pf, which couldn't be reproduced by olivier@. tuexen@ has an idea what is going wrong and will provide a patch to the reporter for testing.D43769 fixes this, but adds some instructions to the code path.
- When using mbuf queueing, some dtrace static probes are missed. Fixed.
- There is a performance regression by just loading the HPTS module. This is fixed.
- There seem to be some BBR users. Try to get them testing RACK.
- It was discussed that it would be a good idea to move the freebsd stack from ticks to micro seconds.
rscheff@ is improving AccECN code: D36303. rrs@ will review it and implement it for the rack stack. Also the interaction between LRO and AccECN++ has to be taken into account. D42563 further adds the support of ACKofACK.
D23230 is implementing ECN++. rrs@ will review it.
tuexen@ wants to write a tool which dumps the BBLog information of a TCP endpoint from a life system or a core. Using kvm_read() has drawbacks as glebius@ pointed out, using a python kgdb script may be an alternative. An alternative is to write a python kgdb script which pumps data into tcplog_dumper. This would allow to minimize code duplication. glebius@ will try to work on this soon. thj@ has an initial implementation at tcplog_dumper.
The handling on ECN with TSO by NICs was discussed. tuexen@ will add a configuration option for various Intel NICs to make the TSO behavior configurable, rscheff@ will make in the TCP stack configurable wether classical ECN or AccECN will use TSO. tuexen@ wants feedback on D44258 and D44259.
- rgrimes@ reports that FreeBSD using the default parameters can not saturate a Gigabit pipe with a delay used by recent access technologies. He will test some parameter changes and report back.
glebius@ is improving the UNIX domain socket implementation as part of his work to improve the generic send and receive functions. He wants feedback on D44151.
glebius@ will re-activate D35199 which speeds up finding bugs.
- glebius@ asked about disabling MSG_OOB for TCP sockets. It would simplify the socket layer, but is not deprecated by the IETF.
- tuexen@ will add fast open tests to the FreeBSD testsuite. This might require some additional counters.
- It was discussed how network related performance testing could be done...
Next Meeting
18 April 2024 @ 1500 UTC using FreeBSDTCPCall