Running folding at home on thread ripper

[Glynn1977]

Running folding at home on thread ripper 24 cores cpu how many work units should i run or can i run and get good results can i run too many and have it not work as fast on each work unit slots

[kiore]

Running folding at home on thread ripper 24 cores cpu how many work units should i run or can i run and get good results can i run too many and have it not work as fast on each work unit slots

Hi just run one instance using -1 in advanced control (default all resources) or set to the number of cores you wish to use. Threadripper should fold pretty well but if you have a GPU running as well you may need to reduce CPU resources to feed the GPU in order to maximise output.
To get useful advice on maximization we will need to know what OS and whether you also running an video cards as well.

[jrweiss]

Running 24 cores at -1 leaves 23 cores, a prime number that is not liked by F@H. Set it for 22 cores if you are doing nothing else with the computer. Monitor your CPU temperature. If your cooling is weak, you may have to reduce further to 20 or 16 or 12...

[KtC]

I would run tests and take some values. In my case 24 threads are normally limited to 22 cause every GPU gets one thread and next is limited to 21 threads due to decomposition. Next test would be to create two slots 11 threads per each. In my particular setup the performance is better when I run one big slot. I think this may be related to additional overhead for downloading/uploading WUs for two slots.

Monitor your CPU temperature. If your cooling is weak

Very true! In case of Ryzer arch the total power of the CPU may be downgraded with PBO settings as well if coooling solution is not very stable.

[bruce]

22 isn't a good choice due to the prime factor 11. I'd find a better way to split it up. 12 and 16 and 18 are always good choices as well as smaller numbers made up of the factors 2 and 3.

[Neil-B]

An 18 cpu slot probably best ... leave the other 6 to cover the 2 gpu slots and have a big to spare ... some of the current projects appear to load up the cpu a bit so the spare is probably a good thing ... wouldn't bother with a second small cpu slot - leave the spare to ensure gpus std properly serviced

[bikeaddict]

Not sure why people are talking about cores instead of threads. ThreadRipper has 48 threads. On Windows, there is a 32-thread limit.

In FAHControl, I'd recommend leaving the slot's thread count set to -1 and see how many threads it uses in the Log tab of FAHControl or in Task Manager or the monitoring app of your choice. If there are spare threads left, you can either add another CPU slot with a thread count set to the spare count, or you can run Rosetta@Home or OpenPandemics under BOINC and limit the CPU usage to a percent that is the equivalent of the spare threads.

[Neil-B]

tbh had taken the OPs original post to mean SMT was turned off ... but I run intel so maybe that isn't something threadripper users do ,,, if running 48 threads and 2 gpus on windows then a 32 thread cpu slot and a 12 thread cpu slot may be best fit if not pushing too hard for thermals?

[bruce]

Not sure why people are talking about cores instead of threads. ThreadRipper has 48 threads. On Windows, there is a 32-thread limit.

if you're going to run on GPUs, you need to allocate at least one thread per GPU and it doesn't hurt to leave a apare or two.

If you're going to fold with your CPU, FAH gets significantly more performance out of a core than a thread. Remember a pair of threads compete with each other for the shared resources of a FPU.

[bikeaddict]

Interesting that the F@H FAQ recommends keeping it on.

https://foldingathome.org/faqs/smp/dete ... threading/
"Hyperthreading is usually enabled in the BIOS by default, and we recommend that it stays enabled, as the SMP cores can use it to process Work Units faster."

But maybe it's not universally better on all the hundreds of Intel and AMD CPUs that have been released over many years.

Paragon will hopefully have an answer one of these days in Part 2.

https://greenfoldingathome.com/2020/05/ ... f-threads/

[Neil-B]

I use ht on my xeons ... but in no way do I get double throughput compared to without ... I dont mind the lower ppd/watt and greater space warming so I go for this but it isn't an easy call for some ... it also varies a bit dependent on project/wu and gear/thermal setups - ht/smt runs hotter and can cause throttling or lower boosts where cooling is an issue

[victor_pp]

There's a subtle difference here: I'd also recommend to "keep it on", that is, do not deactivate it in the BIOS and keep it available. But when you actually run work units for Folding@home or other computation, use as many processes as you have real cores (not threads). Usually that will result in a better result (faster computation times and/or lower temperature / lower fan speeds) than if you put one process on each thread. Also, you can keep a bit of hyperthreading for your desktop needs, if you use that computer as a desktop. But please, do test for yourself, you might be the exception from the rule

[lafrad]

On my AMD and Intel systems, having FAH dive into the Hyperthreading concurrency has *always* helped. There is diminishing returns... my 4 core 8 thread CPUs had some boost from 4->6, but nothing really from 6->7->8 (with 8 being slower than 6, but still faster than 4).

I had a few threadripper experiences, and I'm finding that using about 70-80% of the total "threads" is the best PPD, with a few less than that to save on electricity/heat during the hotter summer months.

I haven't been keeping track on the DB, but its what I see from immediate PPD observations, and overall trends reported on the overclockers.net tracking site.

Keep in mind, many of the TR's have multiple "Chiplets", and it may be more important to have the FAH "Cores" split with friendly CPU counts in them along with some process management to keep those assigned to the right chiplets to get maximum efficiency. (I use process lasso https://bitsum.com/ to keep 2 FAH cores assigned to 2 chiplets each (3970x), and its been quite nice in keeping the system snappy and not bogged down with task switching, so it can run nearly all the time)

In the end, I'd just leave hyperthreading on, assign ~75% of the threads in the system to do FAH work when you aren't doing other things, and get your GPU going strong. The GPU is still significantly more powerful than the processors, and you'll be scraping for points on the CPU when the GPUs just churn through the work.

[KtC]

Keep in mind, many of the TR's have multiple "Chiplets"

The good quality memory is probably much more important for performance than disabling hyper threading.
In case of 3900x (two dies) I observed optimal performance when memory freq is selected according to IF limitations. At least in synthetic benchmarks.

22 isn't a good choice due to the prime factor 11. I'd find a better way to split it up. 12 and 16 and 18

In fact FAH automatically reduces the threads to 21 from 22. I don't think it can be forced to 22. Never tried 16 and 18 setup.

[MeeLee]

Threadrippers, or systems with more than 10 threads, don't need an additional thread for Windows.
If your threadripper is 24 threads, you can run 24 threads, and the OS will still work fine.
This is because each thread has a small amount of headroom.
All the headroom combined (of 10+ cores or threads) is enough for your OS to run comfortably.

You could also try to tune it by looking at your Taskmanager.
Sometimes running 22 cores still shows 100% CPU utilization (in a case where each thread is over-committed).

There's a third method, which some people suggested, in case your energy bill is too high,or your CPU runs too hot.
Is simply to disable SMT. This will free up some resources, in exchange for higher performing threads.
In many cases, running more threads at a lower speed, is preferred over running fewer threads at a higher speed.
But the latter can lower case temps, and power draw by a bit.

Lastly, the amount of GPUs that are folding.
It's wise to set 1 core aside per GPU, as the GPUs usually offer much higher PPD rates.
If you run 1 or 2 GPUs on your threadripper, you can assign 22 threads to the CPU, and won't need an additional thread for Windows.