FahCore_15.exe CPU Usage Spikes

[compdewd]

I currently run my primary system at SMP:6 out of 8. I run Process Explorer in the background with the CPU graph in the taskbar. I noticed that the graph at intervals will show use of 100% of the CPU rather than 75%. I opened up the window for Process Explorer, watched for a few seconds, and noticed that the culprits were the two "FahCore_15.exe"s. They were jumping from 3.5% CPU usage to 10-11% at certain intervals. With all of the other background programs, the CPU for the single FahCore_a4.exe is pushed down to 60-65% instead of 75% when the spike in CPU usage occurs. Of course, this is negatively impacting the CPU slot.

My question is: why is the CPU usage for FahCore_15.exe not constant (3% or 10%)? Is this a known problem or do I need to start troubleshooting?

[PantherX]

On majority of systems that fold FahCore_15 WUs, the CPU usage is almost insignificant for most projects. However, I read that some projects use slightly more CPU Usage (negatively impacting the SMP) so what projects are currently folding on your GPUs?

[compdewd]

The current projects being worked on are 8070 and 8071.

The interval between spikes is about 29 seconds and the spike lasts for about 8 seconds. TPF for the projects is 3 minutes 14 seconds.

[artoar_11]

The current projects being worked on are 8070 and 8071.

The interval between spikes is about 29 seconds and the spike lasts for about 8 seconds. TPF for the projects is 3 minutes 14 seconds.

Your TPF 03:14 is good.
My GTX460 @ 750 MHz on p8070-8074 TPF: 03:12, -smp 4.

[bruce]

This is not something that's under the control of FAH, so while your questions may be interesting, they're not a FAH bug.

A GPU cannot work without some help from the CPU. The amount of CPU being used is mostly dependent on the drivers and the associated middleware that support the GPU. AMD/NVidia do not report how much CPU time is used by the drivers nor does the Pande Group report what else is needed by the OpenMM code. So the only source of information that I know of comes from reports like yours. Replacing driver version x.xx with y.yy may increase or decrease the CPU requirements, but in most cases won't change it. It also depends on the particular Project being folded.

Stanford benchmarks projects running a single FahCore, not combinations, so the "loss" of productivity in SMP caused by a GPU is never evaluated except by Donors like you. In most cases, the recommended priority/affinity/etc. settings that devote processing first to the GPU core and second to the CPU core produces more PPD that the reverse, but that does not guarantee that there won't be any exceptions since the PG doesn't test those conditions, nor can anybody guarantee that Donors have tested all possible combinations of hardware and all combinations of Projects.

If I had to guess, your reported 29 second interval is probably when the FahCore invokes a JIT compiler -- but I don't have any basis for that conjecture except a guess.