Windows: use more than 32 cores?

[Spurkey]

I have 2 Windows machines that FAH is limiting to 32 cores. Machine 1 is a Xeon W-2195 (18 physical, 36 logical cores), machine 2 is a Threadripper 2990WX (32 physical, 64 logical cores). Both machines are running Win10 Enterprise 64-bit. FAH limits the work it does on both machines to 32 cores, and no adding a second slot doesn't fix it - that just makes 2 CPU slots with 16 cores.

My suspicion is that this happens because the client wrapper is 32-bit. The worker core apps are 64-bit, they're separate entirely from the client, but the bit that partitions out the work appears to be 32-bit - 32-bit apps cannot use more than 32 cores. In the System Info tab under the Build heading the Platform entry is listed as "win32 10" and the Bits entry is "32".

Might it be possible to get a version with a 64-bit Windows client?

[foldy]

Yes you can add a second CPU slot and configure both CPU slots to use 32 threads each.

[Spurkey]

Well poo. If a second slot is added and the CPU count left at -1 (for the Threadripper the thinking was that this would still allocate 32 each) for both it doesn't work; if the cores are explicitly specified for each slot then yes it does work.

The question still remains: why is there no 64-bit client for Windows? The workers are 64-bit. A 64-bit client would remove the need to do this.

[JimboPalmer]

The question still remains: why is there no 64-bit client for Windows? The workers are 64-bit. A 64-bit client would remove the need to do this.

Because a 64 bit client would not solve this proble., (and the F@H programmer would have to coordinate 4 clients instead of 3, slowing down all development)

There are Windows issues using more than 32 cores in a single application, but that is not solved with a new client. It is solved by Enterprise Windows

[bruce]

Have you purchased an Enterprise Windows license or a Home Windows license?

[Joe_H]

Besides what has been mentioned so far about the client being 32-bit, that is the FAHClient portion which does not do the actual calculations. It handles downloading and uploading WU's, and communication for starting and stopping them.

The actual calculations are done in separate processes that will show as FAHCore_nn. Those are already 64-bit. A7 is the CPU folding core, 21 and 22 are the current GPU folding cores.

[Weisskeinen]

And how to use more than 64 cores? My Epyc server running on Windows Server has 96 cores, which are divided equally into two processor groups. I tried four CPU slots with 24 cores each, which runs fine, IF I assign the work manually to the processor groups. If windows does it, everything is put into the processor group of the calling task and half of the server is doing nothing while the other half drowns in work.
Do you have any solution for these cases?

[JTorset]

https://www.anandtech.com/show/15483/am ... x-review/3

[Joe_H]

And how to use more than 64 cores?

Under Windows, not possible on a single CPU slot. The limit is 32 cores assigned to each CPU folding slot, 3 separate slots of 32 would use 96 CPU core threads.

Under Linux you would not run into these limits.

[lafrad]

Huge slots are fun... is there any project that is actually better served by that?

Just create 24 CPU slots until you use up all of your processors.... at least it'll get moving for now, and you'll get around that limitation.

[PantherX]

I have seen some CPU Projects running successfully on 32, and 54/56 CPUs. Generally speaking, there are fewer CPU projects for high CPU counts (24 or higher) than for lower CPU counts (less than 24). Occasionally, you may be assigned a WU that can't run on a high thread count as you will see this error in your log file:
There is no domain decomposition for ## ranks that is compatible with the given box and a minimum cell size of ## nm
You can report it in the Forum so that the researcher can prevent it from being assigned to that CPU number and you can reduce the CPUs assigned to it to ensure that you can successfully fold it.