Trying Out The "Folios" Patches On An AMD Linux Server

[phoronix]

Phoronix: Trying Out The "Folios" Patches On An AMD Linux Server

One of the low-level exciting kernel advancements being worked on at the moment is the new "folios" struct for improving Linux memory management. Tests by those involved found that in some conditions Linux kernel builds for example could be up to 7% faster. Given the recent folios v14 patches being published, I took them for a spin on an AMD EPYC server to see the impact on overall performance...

https://www.phoronix.com/scan.php?page=news_item&px=Folios-v14-Testing-AMD-Linux

[tildearrow]

Hmmm, it appears to boost write performance. Maybe you should do a bit more testing in that regard.

Oh wait, you did a lot of testing but why do I only see 19 tests? Where are the other 99?

[QuImUfu]

Hmmm, it appears to boost write performance. Maybe you should do a bit more testing in that regard.

Oh wait, you did a lot of testing but why do I only see 19 tests? Where are the other 99?

https://openbenchmarking.org/result/2107183-IB-MEMORYFOL07

There is a tiny link to that below the system overview table.

[coder]

Michael, what about compilation performance or browser benchmarks?

[arQon]

That looks almost ideal, frankly: a trivial improvement in the majority of cases, a trivial degradation in a tiny majority, and an overwhelming improvement in page-heavy ones. All of that is in line with expectations based on the original outline (and my experience in related work). Short of some catastrophic bug appearing, this looks good to go.

[CochainComplex]

very promissing results. I'm really looking forward to see this thing mainlined.

[intgr]

A response from Andres Freund (a PostgreSQL developer) about the pgbench read-write test:
https://lwn.net/ml/linux-kernel/2021...3.anarazel.de/

I got about 7% improvement with just some baseline tuning of postgres applied.

The test also only runs for 15 seconds, which likely isn't even enough to "warm up"
(the creation of the data set here will take longer than the run).

Given that the dataset phoronix is using is about ~16GB of data (excluding
WAL), and uses 256 concurrent clients running full tilt, using that limited
postgres settings doesn't end up measuring something particularly interesting
in my opinion.