New Linux Scheduler Patches Can Improve AMD Zen Performance For Some Workloads

[phoronix]

Phoronix: New Linux Scheduler Patches Can Improve AMD Zen Performance For Some Workloads

A set of two patches under review on the kernel mailing list for tweaking some kernel scheduler behavior can provide noticeable performance benefits to those using AMD EPYC and Ryzen processors on various workloads...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-Sched-NUMA-Imbalance-Zen

[tildearrow]

So... Linux was already more optimized than Windows in NUMA cases, but now is even faster? Good job...

[ms178]

Now that AMD CPUs enjoy more market relevance and show up in more and more systems of Linux Kernel developers, they are finally squeezing out more performance out of AMD's chips than they used to before.

[numacross]

Where there isn't a 1:1 relationship between the last-level cache (LLC) and node

This can be forced with a BIOS setting (L3 cache as NUMA) on EPYC at least. Which is what AMD recommends for some workloads.

[kozman]

So... Linux was already more optimized than Windows in NUMA cases, but now is even faster? Good job...

In just reading the LKML info about it, it seems it wasn't exactly low hanging fruit. But it's good that there's more perf to be squeezed out of existing AMD CPUs. How much of this will benefit the average user / distro?

2022 is going to be looking good for AMD if this mainlines into 5.17 with the potential for back-porting to earlier kernels where possible. The numbers sure speak volumes in terms of benefits. Now if more eyes and even smarter folks look at what he's trying to do, this might turn into another io_uring-type perf win as well.

[perpetually high]

I gotta say, I'm really loving my Ryzen 5800X. It's a lot of fun. You just enable PBO, and let it do its thing. I've found that schedutil performs better/more consistent than performance governor. It's also nice that it's meant to hit 90C and you're not supposed to worry about it. (Also, mitigations=on and mitigations=off barely affect performance, which is nice).

I got some benchmarks coming soon. Out-the-box it's lightening fast, and I compiled my tweaked-out kernel the other day, and holy crap. Crazy fast benchmarks. Gonna post em soon.

edit: some preliminary pts/osbench tests: https://openbenchmarking.org/result/...TJ-OSBENCHTE66

too lazy; didn't click:






It's crazy how much performance is _still_ left on the table by compiling your own kernel. A mixture of -march=native -O3 + optimizations/patches. Stay hungry, my friends.

edit: compiling the kernel in all its glory (maxing out at 86c right now with Kryonaut Thermal Grizzly instead of 91c with a cheap Cooler Master thermal compound. Using a Dark Pro 4 cooler (BK022).

https://vimeo.com/manage/videos/652316455/customize

[Grinch]

You just enable PBO, and let it do its thing.

On my 3700x Zen 2 I'm getting ~11% performance increase with PBO when rendering in Blender. Also Zen 3 PBO utilization is likely improved compared to Zen 2. When using PBO I max out at ~81c, using a Cooler Master.

A quick tip: You can enable and disable PBO by using: echo 1 > /sys/devices/system/cpu/cpufreq/boost to enable it and 0 to disable it, thus not having change BIOS settings.

[creative]

perpetually high

I actually saw my R5 3600 getting faster and latency seemed reduced more and more before I got my R7 5800x. I thought to myself, "Why am I upgrading?" Oh yea, that extra headroom and ipc uplift.

At stock on very intensive long renders, we are talking strenuous Kdenlive rendering hitting 99% cpu usage it will temporarily hit 91 for the Tdie and Tccd will at least once or twice hit very briefly 92c-94c. On long steam shader compiles it only does 90c though and it takes a bit to get there. I have a Mugen 5 rev. b in push pull with 3 maglev fans intake in front and one in back of case. None of it bothers me too badly, I still tune thermal limiting to see how close the performance is at different temperature limits. Its hard to tell anywhere from 74c to 90c. Heck even 72c. Some people just put these chips on 65watt eco mode. They are indeed interesting chips.

[perpetually high]

Grinch creative thanks for sharing. There's so much to learn about Zen 3. I love it.

What I'm having trouble understanding is if I'm hitting 4.5GHz on all cores with PBO and I'm around 81-86c, so I have about 4-5 degree threshold to "supercharge" PBO, is that correct? I saw some videos and my motherboard (Gigabyte X570S AORUS Master) has some pretty nice features that can really unleash this chip if this cooling is sufficient. So while I'm not on watercooling, the Dark Rock Pro 4 is a great cooler and I've got thermal grizzly on there, so it should have some room to go.

This Gigabyte board has a feature called "Active OC Tuner" that switches between PBO and manual O/C to maximize both single-core and mult-core performance. I haven't dived too deep into it, but it seemed like Active OC Tuner enabled resulted in slower benchmarks, so probably not that great of support on Linux yet, but I haven't done enough benchmarks to be definitive about it. Still lots to learn, will continue to share my findings.

[perpetually high]

Gonna put it inside the case this weekend.