MySQL Bugs: #100119: Switch to use relaxed memory barrier for redo-log counters

Bug #100119	Switch to use relaxed memory barrier for redo-log counters
Submitted:	6 Jul 2020 7:07	Modified:	14 Aug 2020 13:10
Reporter:	Krunal Bauskar	Email Updates:
Status:	Verified	Impact on me:	None
Category:	MySQL Server: InnoDB storage engine	Severity:	S5 (Performance)
Version:	8.0.20	OS:	Any
Assigned to:		CPU Architecture:	ARM
Tags:	arm, memory-barrier, performance

Description:
Switch to use relaxed memory barrier for redo-log counters

- REDO Log flow needs multiple counters to track stats like redo-log size,
  free-spaces, flushed_upto lsn, checkpoint_upto lsn, etc....

- All of these counters are meant to track the state of redo log and not for
  the so-called synchronization of critical resources.

- This naturally qualifies them to use relaxed memory barriers as proposed
  in the said patch.

How to repeat:
Apply the patch that I plan to submit as contribution and run some basic performance tests. I tried sysbench performance and observed performance gain with patch applied.

patch for the bug

Attachment: perf#logtypes-memory-barrier.patch (text/x-patch), 4.62 KiB.

arm benchmark

Attachment: arm.png (image/png, text), 104.72 KiB.

x86 benchmark

Attachment: x86.png (image/png, text), 96.44 KiB.

Hi Mr. Bauskar,

Thank your for your performance improvement report.

I have tested it and I conclude that it is correct.

Verified as reported.

I see some drops with the patch esp. at 64 threads on Intel. This makes it difficult to accept the patch.

Thank you Sunny.

This is a patch for ARM architectures only .......

x86 re-benchmarked

Attachment: x86_8021.png (image/png, text), 48.86 KiB.

I presume you tried the patch with 8.0.21.

Since I submitted the patch back with 8.0.20 (with performance results for both x86 and ARM) I thought may be something could have changed and so decided to give it a re-try with 8.0.21.

I am attaching the results for x86.
I don't see any regression for 64 threads (neither for higher scalability too).

------------------

It would be really helpful if you can share the complete result-set
along with configuration and test-case if you still see the regression.

I will try to reproduce it.

Thank you, Mr. Bauskar.