Windows Doesn't Like 64 Threads; TR 2990WX Threading Tests

[Taf the Ghost]

https://www.phoronix.com/scan.php?page=article&item=2990wx-linwin-scale&num=1

 

Brought this over here because it's more of a discussion than a News piece, but it's interesting none the less.

 

 

Phoronix did a lot of testing and these types of results are pretty common. @leadeater , looks like we have some answers about what is going on. Some of the Windows libraries are likely the culprit. It also explains why there were similar problems when Threadripper 1 dropped. They just moved those libraries from 24 threads to 32 threads and the problem went away. Then AMD drops a 64 thread part and, suddenly, those SMT problems crop back up.

 

So, if you have an AMD 2990WX (or 2970WX) and you're wondering why certain tasks seem really slow, here's the main answer until Microsoft gets around to fixing things. 

 

Edit: Fixed the image.

[Jurrunio]

Use Windows Server?

[leadeater]

@Taf the Ghost

Dead image link, had similar problems trying to link pictures off Phoronix as well. Gave up and just used snip/screenshot last time lol.

[Taf the Ghost]

1 minute ago, Jurrunio said:

Use Windows Server?

https://www.phoronix.com/vr.php?view=26716

 

Turns out, it's worse in a bunch of tasks. Don't ask me how Microsoft managed that.

[Taf the Ghost]

1 minute ago, leadeater said:

@Taf the Ghost

Dead image link, had similar problems try to link pictures off Phoronix as well. Gave up and just used snip/screenshot last time lol.

Yeah, he's using SVG which is a vector/data approach for web display. I didn't realize it, so hard image is go!

[orbitalbuzzsaw]

Microsoft - Micro performance

[Taf the Ghost]

2 minutes ago, leadeater said:

@Taf the Ghost

Dead image link, had similar problems try to link pictures off Phoronix as well. Gave up and just used snip/screenshot last time lol.

As for the topic, it's good we finally have an explanation. Some of it is the scheduler, I imagine especially if you're dealing with Games, but the results were too weird when some tested with SMT Off. It definitely is some set of libraries that have dependencies that just don't accept 64 threads.

 

Most of the time, the performance at 64 threads is around the 24 thread result, which probably means it is doing some sort of fall-back limitation. Funniest result, though, was FFmpeg.

 

 

I imagine this would repeat itself on even X299 CPUs as well. And I do find it funny that 64 threads is slower than 1 thread.

 

However, I think we need to start asking questions about Adobe Premiere performance over threads.

[PlayStation 2]

8 minutes ago, Taf the Ghost said:

Yeah, he's using SVG which is a vector/data approach for web display. I didn't realize it, so hard image is go!

Here, high-res SVG to PNG conversion.

 

A bit more on-topic: hoo-wee. Linux is the way to go for these chips?

Would make for a hella encoding machine.

[2FA]

Isn't this just a repeat of the existing news thread?

 

[PlayStation 2]

Just now, 2FA said:

Isn't this just a repeat of the existing news thread?

 

Not necessarily; this discusses more that Windows doesn't do well with 64 cores rather than Linux performance on the 2990WX being much better than Windows performance.

[leadeater]

2 minutes ago, Taf the Ghost said:

As for the topic, it's good we finally have an explanation. Some of it is the scheduler, I imagine especially if you're dealing with Games, but the results were too weird when some tested with SMT Off. It definitely is some set of libraries that have dependencies that just don't accept 64 threads.

A lot of it looks like software itself as well, there was a decent number which worked fine.

[Taf the Ghost]

1 minute ago, Dan Castellaneta said:

Here, high-res SVG to PNG conversion.

 

 

A bit more on-topic: hoo-wee. Linux is the way to go for these chips?

Would make for a hella encoding machine.

Thanks, I edited the image in. 

 

As for the 2990WX, with a properly working OS, the memory bandwidth issues seem to be fairly minimal, so, yes, they are monster encoding machines.

Just now, 2FA said:

Isn't this just a repeat of the existing news thread?

 

We knew the Linux performance was a lot better, but this thread is about "why". There's clearly problems with the way Windows handles certain things, and it's not just the Scheduler.

[Taf the Ghost]

1 minute ago, leadeater said:

A lot of it looks like software itself as well, there was a decent number which worked fine.

Yup. It's more than likely some set of standard libraries that has the issues. Maybe a set of compiler flags?

 

Interesting thing is that this should hit the 28 core Intel A-series as well.

[leadeater]

7 minutes ago, Taf the Ghost said:

Yup. It's more than likely some set of standard libraries that has the issues. Maybe a set of compiler flags?

 

Interesting thing is that this should hit the 28 core Intel A-series as well.

Probably a lot of stuff using .Net and/or older stuff with fixed thread pool sizes or no thread pooling at all. From my limited time doing C# programming multi-threading and thread scaling really is not a "it just works" thing. Been a long time since I've written any sort of proper code though.

 

Edit:

https://docs.microsoft.com/en-us/dotnet/standard/parallel-programming/task-parallel-library-tpl

[Taf the Ghost]

8 minutes ago, leadeater said:

Probably a lot of stuff using .Net and/or older stuff with fixed thread pool sizes or no thread pooling at all. From my limited time doing C# programming multi-threading and thread scaling really is not a "it just works" thing. Been a long time since I've written any sort of proper code though.

 

Edit:

https://docs.microsoft.com/en-us/dotnet/standard/parallel-programming/task-parallel-library-tpl

 

And just before this timestamp as well. I think a lot of those libraries break at 32 threads. We haven't really looked at a whole bunch of HEDT benchmarks since last year, and there's been a lot more going on here than we realized. Mostly because they added just so many extra threads in such a short time-span that I don't think anyone realized there were problems since the results were so much higher in raw numbers.