maria-developers team mailing list archive

[Stewart Smith <stewart@xxxxxxxxxxxxxxxx>]

Colin Charles <colin@xxxxxxxxxxx> writes:
> I'm sure you have read:
> 	https://www.flamingspork.com/blog/2014/10/08/mariadb-10-0-on-power/
>
> For XL:
> There’s obviously still some work in progress going on, but my initial
> results show performance within around 10% of MySQL, so with a bit of
> work we will hopefully see MariaDB reach performance parity.

I didn't delve into it too deeply, but am happy to help where I can.

> For Serg:
> One interesting find was the code to account for thread memory usage
> uses a single atomic variable: this does not scale and does end up
> showing up on profiles.
>
> It isn't clear 10% loss to which MySQL, but comparing to 5.6 or 5.7,
> this has to "just work better"

It was 5.6.something (20? I think with one patch from me). Both compiled
the same way.

IIRC it was 2-3% of total execution time spent doing this accounting. I
doubt it's worth it - and it's likely that it can be done more
efficiently too. Bouncing that single cache line around all CPU cores,
especially if mariadb is run on more than a single NUMA node on POWER8
is going to hurt *BAD*.

These kind of issues are actually pretty easy to spot using perf on
POWER, and my suspicion is that the 10% perf difference is made up of a
couple of things like this and shouldn't be tooo hard to claw back.

-- 
Stewart Smith

Attachment: pgpv8dM4XKSr3.pgp
Description: PGP signature