CPU requirement for GPU folding (NV)

[Aurum]

So far I'm not seeing any improvement using Project Lasso but maybe I should assign middle CPUs. For a dedicated folding rig I think if one just leaves enough headroom it'll work fine.

[Joe_H]

The problem of F@H seeing only 32 threads max on a system with more available has been seen before. One example I recall was due to the CPU being an engineering sample, it was recognized as a different type with fewer threads available than it had. But most other times it was related to which OS was being used and the settings within that OS. Windows in particular depends on which release and license is active.

[Aurum]

One example I recall was due to the CPU being an engineering sample, it was recognized as a different type with fewer threads available than it had. But most other times it was related to which OS was being used and the settings within that OS. Windows in particular depends on which release and license is active.

It's an ES version.

But most other times it was related to which OS was being used and the settings within that OS. Windows in particular depends on which release and license is active.

Win7 Task Manager & CPU-Z saw all 44 threads before I rebooted and disabled hyperthreading. BTW, this rig is working great sans hyperthreading. Of it may just be a fast CPU. I think I'll do that on more rigs but since they're headless it'll be a bit of work.

Does F@H use a lookup table for CPUs??? Are you sure this is not a coded limit???

Now have 22c22t and set F@H to use 18 threads for folding. With three 1080 Ti cards (x16x16x8 3.0) that leaves one for other stuff which stays around 20% use. All 22t are at 95-96% use.

[Joe_H]

There are process limits within Windows that apply. With the appropriate versions, licenses, and settings those can be increased to the full number of threads available. So the OS sees all cores including those from HT, but a single process can only use 32 unless your licensed version allows more and the correct settings have been made in the OS.

I am fairly certain this is not a coded limit in F@H, people have tested it on systems with many more cores than you have. The highest I am aware of was on a Xeon Phi with 256 cores available.

[bruce]

The fact that FAHClient/FAHCore are limited to 32 CPUs shouldn't be a problem even if your system has 44 threads. You just need to create a second CPU slot.

Like Joe_h has said, the 32 thread limitation was based on at least one of the OSs that FAH supports -- and it wasn't worth figuring out and enabling specialized exceptions. FAH is designed for At Home computers and there aren't a lot of home computers with more than 32 threads.

[Aurum]

Which will be faster 18 cores or 36 threads???

[Nathan_P]

Which will be faster 18 cores or 36 threads???

On the same cpu with the same work unit 36 threads will be quicker. IIRC the last time this was tested there was anything from a 5 to 25-30% improvement at least.

[Aurum]

Which will be faster 18 cores or 36 threads???

On the same cpu with the same work unit 36 threads will be quicker. IIRC the last time this was tested there was anything from a 5 to 25-30% improvement at least.

But since I cannot get more than 32 threads per WU I have to set up two CPUs with 18 threads each so I guess that means two different WUs. So choice is:
1. Hyperthreading disabled and use 18 cores for one WU, or
2. Hyperthreading enabled and use two F@H CPUs with 18 cores each for two WUs.

[bruce]

Choose #2

If you run 18 even (or odd) numbered threads, it's essentially the same as disabling HT because each thread has the dedicated resources of a CPU. When you start a second WU on the unused threads, it won't be twice as fast because there's contention for the shared resources-- so the total throughput will be between 5% and 25% SLOWER THAN TWICE as fast, but that's still better than just running one WU.

[Aurum]

It's funny how BOINC POGS runs all 44 threads.

[Nathan_P]

If boinc is seeing all 44 threads I don't see why F@H isn't. have you tried reinstalling the client? Also 36 is a multiple of 9 so the client may not allocate work. Other than that ES cpu's work just fine - I've used them on F@H for 6 years now without issue.

[bruce]

FAH is capable of running as many threads as you have, however he GROMACS analysis program will crash frequently if the number contains certain prime factors such as 11. FAH doesn't wnt to assign simulations that are going to crash before they're completed -- and our donors complain that their PPD goes down, hence WUs will not be assigned to 11, 22, 33, 44 ... and many other combinations that contain other prime factors.

GROMACS is a respected method of simulating protein folding (and BOINC doesn't use it). The code is open source and if you can contribute a code improvement which will eliminate those crashes, it will be a big help to everyone who simulates protein motions. Many experts have fought this problem and the only solution that has been found is to only simulate on systems where the number of threads contains only the factors 2 and 3 (and in some cases, 5)

[Aurum]

I took Bruce's advice, enabled hyperthreading on the 44t and allocated it to 3 bins: 6 for three 1080 Ti, a CPU with 18 & a CPU with 20. The System Info tab still says there's only 32 CPUs but they all run according to Win7 Task Manager. It works great.
I have 2 other rigs set to 22 CPUs and they seem to run fine. I'm trying 10+12.