cxgbe netmap: Micro optimizations of the rxsync function
AcceptedPublic
Actions

Authored by bwicht_verisign.com on Nov 6 2018, 12:16 PM.

Details

Reviewers

np
vmaffione

Summary

Modifying slot->flags only when needed improves by ~10% max receive packet/sec
Remove assertions in the hot path
Improve branch prediction

Submitted by: Marc De La Gueronniere <mdelagueronniere@verisign.com>
MFC after: 1 week
Sponsored by: Verisign, Inc.

Diff Detail

Lint

Lint Passed

Unit

No Test Coverage

Build Status

Buildable 20680
Build 20095: arc lint + arc unit

Event Timeline

bwicht_verisign.com created this revision.Nov 6 2018, 12:16 PM

Herald added a subscriber: ae. · View Herald TranscriptNov 6 2018, 12:16 PM

Harbormaster completed remote builds in B20658: Diff 50069.Nov 6 2018, 12:16 PM

bwicht_verisign.com edited the summary of this revision. (Show Details)Nov 6 2018, 12:17 PM

bwicht_verisign.com added a reviewer: np.

+Vincenzo for the netmap_kern.h part.

np added inline comments.Nov 6 2018, 6:00 PM

sys/dev/cxgbe/t4_netmap.c
915	The assertions shouldn't affect non-debug kernels. I think this one should not be removed.
916	Is this prediction applicable across the board or only for a certain type of workload?

jch added a subscriber: jch.Nov 6 2018, 9:29 PM

Restore removed assertions

Harbormaster completed remote builds in B20680: Diff 50111.Nov 7 2018, 9:59 AM

I have restored the assertions.

sys/dev/cxgbe/t4_netmap.c
915	You are right. I restored the assertions. We have been a bit too far in the change.
916	We are confident that this change will not degrade performance on any load. Our 10% improvement was seen with a very-high load of receive traffic with small packets.

The changes to netmap_kern.h look very good, thanks.

sys/dev/cxgbe/t4_netmap.c
916	This looks interesting. We were wondering wether the improvement comes from the __predict_false or from skipping the memory write operation. This information could be useful in other parts of netmap. Have you tried the two changes independently?

This revision is now accepted and ready to land.Nov 7 2018, 10:28 AM

bwicht_verisign.com marked 2 inline comments as done.Nov 8 2018, 3:43 PM

bwicht_verisign.com added inline comments.

sys/dev/cxgbe/t4_netmap.c
916	Unfortunately, we have not.

Il giorno gio 8 nov 2018 alle ore 16:43 bwicht_verisign.com (Baptiste
Wicht) <phabric-noreply@freebsd.org> ha scritto:

bwicht_verisign.com marked 2 inline comments as done.
bwicht_verisign.com added inline comments.

INLINE COMMENTS

v.maffione_gmail.com wrote in t4_netmap.c:913
This looks interesting. We were wondering wether the improvement comes

from the __predict_false or from skipping the memory write operation. This
information could be useful in other parts of netmap.

Have you tried the two changes independently?

Unfortunately, we have not.

Ok, it does not matter.

CHANGES SINCE LAST ACTION
https://reviews.freebsd.org/D17868/new/
REVISION DETAIL
https://reviews.freebsd.org/D17868
EMAIL PREFERENCES
https://reviews.freebsd.org/settings/panel/emailpreferences/
To: bwicht_verisign.com, np, v.maffione_gmail.com
Cc: jch, ae