Super slow folding - 16 cores = ~500 ppd???

ChrisITguy · November 28, 2020

So I'm folding on some extra hardware I have laying around and one of the systems is an Opteron 6276 (16 cores, 2.3GHz). Since it's winter, any heat this thing generates while folding is fine (house is electric heat anyhow, same cost per kWh lol) Server chassis, completely adequate cooling. Folding@home CPU slot set to -1 (client determines cores), client sets CPU:16, task manager shows totally maxed out (100% across all 16 logical processors) and I'm screaming along at.....527 ppd? For reference, our HTPC is set to use only 1 core, no GPU, and it's doing ~4500 ppd. Office PC is doing nearly 1 million ppd (full folding performance while I'm not there)

Did F@H add something to work units that may be using a feature or extension that these 6276's don't support now? It used to be a lot better than that last time I ran FAH on this particular piece of hardware.

LazyDev · November 28, 2020

6 hours ago, ChrisITguy said:

So I'm folding on some extra hardware I have laying around and one of the systems is an Opteron 6276 (16 cores, 2.3GHz). Since it's winter, any heat this thing generates while folding is fine (house is electric heat anyhow, same cost per kWh lol) Server chassis, completely adequate cooling. Folding@home CPU slot set to -1 (client determines cores), client sets CPU:16, task manager shows totally maxed out (100% across all 16 logical processors) and I'm screaming along at.....527 ppd? For reference, our HTPC is set to use only 1 core, no GPU, and it's doing ~4500 ppd. Office PC is doing nearly 1 million ppd (full folding performance while I'm not there)

Did F@H add something to work units that may be using a feature or extension that these 6276's don't support now? It used to be a lot better than that last time I ran FAH on this particular piece of hardware.

That Opteron 6276 is basically obsolete, unfortunately. Technology advances pretty fast these days. My Ryzen 3700X can get over 100K PPD using all 16 threads at once. This is due to many variables that have improved over time. Just three of my CPU threads beats your entire 16 core CPU, for example.

For comparison, here's my CPU compared to yours, https://www.cpubenchmark.net/compare/AMD-Opteron-6276-vs-AMD-Ryzen-7-3700X/1880vs3485

Your only option is to upgrade, or to stick a GPU into the system.

ChrisITguy · November 28, 2020

Doesn't really answer the question though as why it went from a much higher ppd down to 500 ppd or why a single thread on a haswell i7 is now 8x faster. And if your Ryzen is at 100k ppd, that is 200x faster. Your CPU isn't 200x faster in any benchmark . Something is still amiss...

rkv_2401 · November 28, 2020

Have you entered your passkey for the Opteron client? Hardware wise, there's no way a 16 core Opteron should generate just 500PPD. I had an Athlon 64 3200+ that was putting out 500PPD back in 2006-7. From what I can tell, quad core Sandy/Ivy Bridge CPUs should be getting around 20-25k PPD, so yours should be doing much higher than that.

(https://foldingforum.org/viewtopic.php?f=88&t=2601,

http://www.prowaretech.com/Computer/It/FoldingAtHomePPD )

EDIT :

Here's a thread on Opteron's F@H performance :

https://hardforum.com/threads/improve-opteron-6200s-performance-and-ppd-by-using-turionpowercontrol.1708870/

ChrisITguy · November 28, 2020

Ended up getting it sorted out myself - I remembered reading somwehere that the 16-core Opterons were actually a 16-core chip with 8 FPUs (2 cores share an FPU). I decided to try forcing the cores slot to 8 (instead of -1 / auto), client is using 8 CPUs now and ppd went from ~520 to ~240k ppd. All I can guess is that having 16 threads competing for 8 FPUs was causing too much contention for a FPU slot and jamming up thread handling. System shows ~50% CPU load now which is fine if it turns out more ppd. If I go up any higher on the CPUs, ppd actually drops. To test the theory, I went up to 10 CPUs and ppd drops to 8.9k ( !!!! ). Dropping it to 8 cores jumps it right back up.

I don't remember having this issue when I ran FAH on it before but who knows. Maybe something changed with the client or the work units.

Thanks for giving it some thought though!

Cheers,

Chris