tcp: update "cubic_k" with the new formula from RFC8312bis.
Needs RevisionPublic
Actions

Authored by rscheff on Aug 31 2023, 9:29 AM.

Details

Reviewers

cc
tuexen
jtl
rrs
glebius
lstewart

Group Reviewers

transport

Summary

2nd patch in series to update Cubic in conformation with RFC8312bis.

This updates the calculation of K using the formula from this RFC document.

The new formula additionally uses "cwnd_epoch" for calculating K.

(1st in series: D40436)

This Diff is sliced from a summary patch provided by VMware to improve short RTT performance of Cubic.
Once all patches have been prepared and reviewed, they will go in together to make an atomic change
to cubic from pre-RFC status, to 8312bis status.

Test Plan

Submitted By: Bhaskar Pardeshi, VMware Inc.

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Passed

Unit

No Test Coverage

Build Status

Buildable 53369
Build 50260: arc lint + arc unit

Event Timeline

rscheff created this revision.Aug 31 2023, 9:29 AM

Herald added a reviewer: transport. · View Herald TranscriptAug 31 2023, 9:29 AM

Herald added subscribers: melifaro, imp. · View Herald Transcript

rscheff requested review of this revision.Aug 31 2023, 9:29 AM

Harbormaster completed remote builds in B53364: Diff 126724.Aug 31 2023, 9:29 AM

fix math typo

Harbormaster completed remote builds in B53369: Diff 126736.Aug 31 2023, 2:35 PM

another typo

Harbormaster completed remote builds in B53370: Diff 126737.Aug 31 2023, 2:39 PM

another typo

Harbormaster completed remote builds in B53371: Diff 126738.Aug 31 2023, 2:44 PM

rscheff mentioned this in D41672: tcp: add cubic_compute_pipe.Aug 31 2023, 2:48 PM

cc requested changes to this revision.Sep 1 2023, 9:39 PM

cc added inline comments.

sys/netinet/cc/cc_cubic.c
334	Here, we are mixing the usage of "CUBICFLAG_CONG_EVENT" flag with the usage of "num_cong_events" variable below in the cases of cubic_cong_signal(). We shall not use both. Suggest remove the usage of "num_cong_events" as in commit 6907bbae187e

This revision now requires changes to proceed.Sep 1 2023, 9:39 PM

rscheff added inline comments.Sep 1 2023, 9:46 PM

sys/netinet/cc/cc_cubic.c
334	Yes; I'm not done yet in dissecting the summary patch into smaller snippets; the flags will all be removed and only the num_cong_events remaining... If you think it is unneccessary to split up the patch into logical subunits - which compile but may contain logic shortcomings, I can simply discard these (currently 2, probably will be 6-9) diffs and provide a single summary Diff?

cc added inline comments.Sep 2 2023, 2:03 AM

sys/netinet/cc/cc_cubic.c
334	Reading probably 6 to 9 patches and looking up connecting logics may also be a challenge. Also test the single patch by emulation is easier. I prefer to a single patch on top of the 1st patch which is committed as eb5bfdd06565. Others thoughts?