Benchmarking The AMD EPYC Speed Boost Coming To Linux 5.18, Thanks To Scheduler/NUMA Improvement

[phoronix]

Phoronix: Benchmarking The AMD EPYC Speed Boost Coming To Linux 5.18, Thanks To Scheduler/NUMA Improvement

Earlier this month I noted a Linux scheduler change queued into sched/core ahead of the Linux 5.18 cycle that is expected to help AMD EPYC processors and other select Zen processors in various workloads. The change has been in the works for several months and is about adjusting the allowed NUMA imbalance when spanning multiple LLCs. I've now carried out some of my own benchmarks on EPYC hardware and indeed is further ratcheting up the Linux kernel performance.

https://www.phoronix.com/vr.php?view=30941

[numacross]

This looks great for people who let their EPYCs run in default configuration due to not reading AMD's optimization documentation

Michael Can you test with NPS4 and L3 as NUMA enabled?

[chuckula]

This looks great for people who let their EPYCs run in default configuration due to not reading AMD's optimization documentation

Michael Can you test with NPS4 and L3 as NUMA enabled?

Then the question is: Why is AMD's default configuration so sub-optimal?

[numacross]

Then the question is: Why is AMD's default configuration so sub-optimal?

Because people are scared of NUMA and rightfully so - for some workloads it is beneficial while for others it's detrimental.
Intel has similar configuration modes available. 3rd generation Xeon Scalable can be split into 2 or 4 NUMA nodes if needed. There is also a special mode for L3, but I have no experience with it so it might not work the same as L3 as NUMA on AMD.

[schmidtbag]

That's enough of a performance difference to be considered a next-gen CPU, or tapping into a new instruction set. Pretty impressive results for a scheduler. Makes sense though.

[kozman]

Not bad for some minor tweaking of the NUMA code! In most tests it's a decent win so that's nice. Always goes to show there's more perf to be squeezed out of AMD CPUS with some thought and effort. I bet after baking for a while someone is going to find a further way to tweak the NUMA stuff after others have had some time to put it into production and beat on it for a while. Nice job with this.

[Beherit]

I wonder what the performance is with NUMA turned off.

For some reason I cannot remember any longer, I turned NUMA off when I got a Ryzen setup.

[kozman]

I wonder what the performance is with NUMA turned off.

For some reason I cannot remember any longer, I turned NUMA off when I got a Ryzen setup.

Would be curious to see this too. Something like NUMA off vs NUMA vanilla vs new NUMA perf code. Big data benefits most from this though, no?

[mlau]

Could you please test a 2-die Ryzen 5000 as well please (5950 for example) ?

[FLHerne]

This shows just how foolish AMD's underinvestment in the kernel (and software generally) has been.

Generational performance improvement from one patch - that isn't even by an AMD employee - years after Rome launched. How many million man-hours did it take to get the 15-20% perf improvement in Milan, while low-hanging fruit like this was sitting around on the software side?

It's been improving for the last year or two, but still a handful of kernel and graphics engineers could do much more to make the product attractive than the diminishing-returns end of hardware updates.