Re: Possible performance regression in version 10.1 with pgbench read-write tests.

On Fri, Jul 20, 2018 at 10:52 AM, Thomas Munro <
thomas.munro@enterprisedb.com> wrote:

> On Fri, Jul 20, 2018 at 7:56 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >
> > It's not *that* noticeable, as I failed to demonstrate any performance
> > difference before committing the patch.  I think some more investigation
> > is warranted to find out why some other people are getting different
> > results
> Maybe false sharing is a factor, since sizeof(sem_t) is 32 bytes on
> Linux/amd64 and we're probably hitting elements clustered at one end
> of the array?  Let's see... I tried sticking padding into
> PGSemaphoreData and I got ~8% more TPS (72 client on multi socket
> box, pgbench scale 100, only running for a minute but otherwise the
> same settings that Mithun showed).
>
> --- a/src/backend/port/posix_sema.c
> +++ b/src/backend/port/posix_sema.c
> @@ -45,6 +45,7 @@
>  typedef struct PGSemaphoreData
>  {
>         sem_t           pgsem;
> +       char            padding[PG_CACHE_LINE_SIZE - sizeof(sem_t)];
>  } PGSemaphoreData;
>
> That's probably not the right idiom and my tests probably weren't long
> enough, but there seems to be some effect here.
>

I did a quick test applying the patch with same settings as initial mail I
have reported  (On postgresql 10 latest code)
72 clients

CASE 1:
Without Patch : TPS 29269.823540

With Patch : TPS 36005.544960.    --- 23% jump

Just Disabling using unnamed POSIX semaphores: TPS 34481.207959

So it seems that is the issue as the test is being run on 8 node numa
machine.
I also came across a presentation [1] : slide 20 which says one of those
futex architecture is bad for NUMA machine. I am not sure the new fix for
same is included as part of Linux version 3.10.0-693.5.2.el7.x86_64 which
is on my test machine.


[1] https://www.slideshare.net/davidlohr/futex-scaling-for-multicore-systems


-- 
Thanks and Regards
Mithun C Y
EnterpriseDB: http://www.enterprisedb.com