MySQL Bugs: #99556: Avoid sequentially consistent atomics for atomic counters

Bug #99556	Avoid sequentially consistent atomics for atomic counters
Submitted:	13 May 2020 18:18	Modified:	15 May 2020 11:48
Reporter:	Sergey Glushchenko	Email Updates:
Status:	Verified	Impact on me:	None
Category:	MySQL Server: Compiling	Severity:	S5 (Performance)
Version:	8.0.20	OS:	Any
Assigned to:		CPU Architecture:	ARM

Description:
std::atomic defaults to sequentially consistent memory ordering.
However there are counters that don't really need to use SEQ_CST
memory order, they only need atomicity. While there is no
difference on x86 CPUs, on ARM servers we can see significant
scalability improvements when using weaker memory ordering.

Good example of such counter is the number of running threads,
which we can get rid of entirely as discussed in bug #99412.

Another example is atomic_global_query_id, which becomes the
bottleneck after fixing the bug #99412.

How to repeat:
Run in-memory sysbench oltp_ps with 1024 threads on ARM server and analyze scalability bottlenecks.

Suggested fix:
The patch below

suggested fix

Attachment: bug99556.patch (application/octet-stream, text), 8.39 KiB.

Why introduce another structure when existing invocations that are limited could be fixed to using fetch_add/fetch_sub with the needed memory model.

HI Mr. Glushchenko and Mr. Bauskar,

Thank you for your important contribution.

However, as far as I understand, your patch requires additional simplification.

Please, do let me know when new patch is ready and I will verify this bug report ASAP.

Hi Mr. Milivojevic,

I think this patch is good to go. Atomics are often used just as counters in server code and InnoDB, their use is not limited to the cases fixed in this patch. In the future I plan to submit another patch fixing some hot InnoDB counters.

Also, introducing the new datatype makes the patch smaller.

And finally, if we need to use one of the counters in the new code or modify existing code we don't need to remember to explicitly specify memory order.

Hello  Mr. Glushchenko,

This report is verified as reported.

Thank you for your contribution.