cc_cubic: Use units of micro seconds (usecs) instead of ticks in rtt.
ClosedPublic
Actions

Authored by cc on May 20 2023, 3:22 PM.

Details

Reviewers

rscheff
tuexen

Group Reviewers

transport

Summary

This improves TCP friendly cwnd in cases of low latency high drop rate
networks.
Tests show +42% and +37% better performance in 1Gpbs and 10Gbps cases.

test result in detail: testD40185

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Passed

Unit

No Test Coverage

Build Status

Buildable 51581
Build 48472: arc lint + arc unit

Event Timeline

cc created this revision.May 20 2023, 3:22 PM

Herald added a reviewer: transport. · View Herald TranscriptMay 20 2023, 3:22 PM

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

cc requested review of this revision.May 20 2023, 3:22 PM

Harbormaster completed remote builds in B51581: Diff 122166.May 20 2023, 3:22 PM

cc edited the summary of this revision. (Show Details)May 20 2023, 3:42 PM

cc edited the summary of this revision. (Show Details)

tuexen added inline comments.May 20 2023, 8:25 PM

sys/netinet/cc/cc_cubic.h
113	Why isn't this converted?
232	Should there be a space between `*` and `tick`?

cc added inline comments.May 21 2023, 12:21 PM

sys/netinet/cc/cc_cubic.h
113	I am not sure if there is a high resolution timer to use for the base stack. So this is a workaround.

correct the lower bound for min_rtt_usecs and correct some space and comments

Harbormaster completed remote builds in B51714: Diff 122486.May 26 2023, 5:44 PM

You might want to wait also for Richards' approval, since he is much more familiar with the code than I am.

This revision is now accepted and ready to land.May 26 2023, 5:48 PM

cc added inline comments.May 26 2023, 5:50 PM

sys/netinet/cc/cc_cubic.h
113	Secondly, cubic_data->t_last_cong as an intermediate variable is used for usecs_since_cong calculation in cubic_ack_received(). The usecs_since_cong is converted by using usec there. So I don't think we need to convert cubic_data->t_last_cong by TICKS_2_USEC(ticks).

In D40185#917355, @tuexen wrote:

You might want to wait also for Richards' approval, since he is much more familiar with the code than I am.

Sounds good to me. :)

Not changing the generic rtt base stack unit / measuring, just improviing the granularity/precision of the calculations within Cubic. since this improves cubic performance, let's put this in prior 14 branching...

This patch is submitted in a3aa6f652904