Folding using only half of dual Xeon Gold 5120's?

[arcangel13579]

Hello,

 

I have what might be a silly question, but I recently installed Folding at home on a machine I have with 2x Xeon Gold 5120's just because it was sitting around, the event is coming up, and I wanted to see if having that much CPU compute would be at all competitive to a GPU points-wise, but what I am seeing is that Folding even with "folding power" set to "full" or me adding multiple CPU slots, it will never use more than about 60% of the CPU, even when manually set to use 55 threads per slot!

 

I was wondering if anyone might have some insight or wisdom into what might be going on, or if there was some other way to increase my folding "density" on this machine to use the wasted CPU time. Maybe the client is NUMA aware? That's about all I can think of...

 

Thank you in advance, and I appreciate it.

 

Some additional info in case it is relevant:

• ProLiant BL460c Gen10
• Server Power Management is set to "Static High Performance Mode"
• Thermals seem normal-ish, CPU is around 55C

 

[Gorgon]

1. You are limited to a max of 32-threads per slot I believe.

2. You can run multiple CPU slots on a Processor

3. Don't let the slots span NUMA boundaries (i.e. chips or chiplets) or performance will suffer. Process Lasso can be useful for this.

4. There is a weird bug where slots can not use certain thread counts (something to do with prime numbers) but will drop down to the next lower safe multiple so expect weirdness.

 

@leadeater will likely be able to give good advice on the best settings.

[leadeater]

Aughh been a while since I had to deal with this issue, about a year (last event haha).

 

The long and short of it is you need to create multiple CPU slots of around 16 each. For you either 2 slots of 14 or 4 slots of 14, which ever actually gives the best overall PPD.

 

I think I have a server with Windows already installed that I can put F@H on and get the settings if you can't resolve it.

 

The easier method is actually ESXi and the F@H OVA appliance or enable Hyper-V role and create a few VMs.

 

12 hours ago, arcangel13579 said:

it will never use more than about 60% of the CPU, even when manually set to use 55 threads per slot!

As @Gorgon mentioned F@H client and projects have a hard limit of 32 threads, there isn't a way to make it use more. The above create multiple slots is the solution.

 

~60% of 56 is 32.

[Gorgon]

1 hour ago, leadeater said:

Aughh been a while since I had to deal with this issue, about a year (last event haha).

 

The long and short of it is you need to create multiple CPU slots of around 16 each. For you either 2 slots of 14 or 4 slots of 14, which ever actually gives the best overall PPD.

 

I think I have a server with Windows already installed that I can put F@H on and get the settings if you can't resolve it.

 

The easier method is actually ESXi and the F@H OVA appliance or enable Hyper-V role and create a few VMs.

 

As @Gorgon mentioned F@H client and projects have a hard limit of 32 threads, there isn't a way to make it use more. The above create multiple slots is the solution.

 

~60% of 56 is 32.

And now is the time to get the tinkering done before the event starts!

[arcangel13579]

Excellent! Thanks for all the advice! Some adjustments to the slots as mentioned does help

 

I could certainly load ESXi and run the appliance as well... if I have time I'll explore that as well.

[leadeater]

1 hour ago, arcangel13579 said:

I could certainly load ESXi and run the appliance as well... if I have time I'll explore that as well.

Multiple slots under Windows I believe would be slightly better performance. I haven't actually benchmarked between the two but the difference should be very minimal. If you can get it using 100% under Windows then I would say leave it as that.