GPU Only CPU running at 30%

[Yavanius]

After digging around to find out to disable CPU processing, I figured out how to go GPU only. Apparently there is no friendly switch to do so.


I'm showing 30% CPU usage though this is suppose to be a GPU crunching. The Advanced client isn't up and there's no visualizations going.

That's another issue. The viewer won't even show me a model. This system does have a primary Intel HD display with the nVidia for the heavy lifting. Two things happened in about the timespan. I deleted the CPU slot and F@H crashed when I was trying to restart it after I had shut it down.

I've tried shutting F@H and restarting with no change.

Thanks.

[PS3EdOlkkola]

What's your question?

[JimboPalmer]

One CPU is always feeding the GPU with data, if you have 4, that would be at least 25%

[Yavanius]

One CPU is always feeding the GPU with data, if you have 4, that would be at least 25%

So it's not quite a pure GPU (as close as one can anyways) client. I'm used to BOINC projects. I didn't realize the CPU use was going to be that high. I wanted to keep BOINC on the CPUs and Folding on the GPU. This screws up the balance somewhat. :/

Thanks for the info.

[Joe_H]

Part of the CPU usage may be related to the driver for the GPU card. The way nVidia implemented its support of OpenCL leaves the CPU core in a spin wait to handle I/O to and from the GPU for the WU data being processed. Then add sanity checks and checkpoints written very few percent of progress. AMD-ATI's GPU drivers handles the OpenCL data transfer with less CPU involvement.

[Yavanius]

Thanks for the info. I believe most of the BOINC projects have CUDA just for nVidia even when they support Radeon too.

Everytime we look like we're getting standardized...

[Joe_H]

PG chose to focus on only having to support one set of OpenCL code instead of maintaining two separate codebases for both OpeCL and Cuda. At one time they did mention that if the Cuda performance was that much higher, and some missing features were added, then a separate Cuda core might be considered.

[Yavanius]

I get that. I don't know if that applies to the BOINC projects or not. It's up to each project to decide and I haven't delved into the history of their GPU developments. Even though I have an nVidia GPU, I wish they had gone the OpenCL route so it wouldn't matter nVidia or AMD/Radeon. :/

Universe@home seems to be the exception in developing on the AMD side with the reason given being that there's better support IIRC. However, they are still testing internally. GPU development is a different beast it seems. It probably was here too.

[bruce]

The fact is, if BOINC + FAH + foreground activities leaves less than one full CPU to drive the GPU, performance will suffer.

There's no such thing as pure GPU processing. Data has to be processed in RAM and submitted to the GPU (a co-processor), the results retrieved, and in RAM data updated. VRAM acts as a scratch-pad memory for that portion of the protein that has been most recently uploaded to the GPU for it to work on.

Think of the GPU as if it's a supercomputer on the cloud and the PCIe bus is like your internet connection to the cloud. If your CPU gets bogged down and quits submitting work, processing slows down.

Configure BOINC so it can't use (at least) one CPU thread. That's the same reason that the default for a FAH-only installation, the number of CPUs allocated to processing CPU-based WUs is always at least one less than the number of CPUs that you have.

[Yavanius]

I experimented last night and saw Folding was still taking up 25% of the CPU even with BOINC maxed out (all 4 cores at 100%). Interestingly, it seemed BOINC wasn't quite sure about this so none of the cores seemed to be hitting 25%. They were maxing out about 20% and it really about 18-19% consistently.

Folding was at Medium.

It's been a long time since I tested both together so I can't recall if that was the similar behavior. I want to say I had to put it on high on the old F@H client, but of course there was no GPU support back then (PS3 aside). I'll test the CPU side once I reach the milestone I'm approaching on a BOINC project which is rapidly approaching. I haven't done much in the life sciences in general in a whiles although I am on RALPH (the Rosetta beta).

[7im]

Neither program was designed to share with another D.C. Project.

[Yavanius]

Designed no, but some accommodations made. Besides, never tell a donor what it doesn't do as they'll have to go and try and make it work... Any day now I expect one of the projects will have some new feature from some volunteer that'll deliver pizza to them...

[toTOW]

I often run both FAH (GPU) and BOINC (CPU) at the same time on my machines ... to let everyone work at full capabilities on my 8 thread CPUs, I set up BOINC to use 80% of the CPUs (which will use 6 threads in my case) and FAH will be happy with the two free threads.

Any way, even in this configuration, FAH GPU performance will be slightly lower than with no load on the CPU ... so when I have no objectives on BOINC or when I need to limit the heat in the room, I turn completely off the BOINC client so that FAH GPU runs at its maximum capability

[bruce]

Have you run an extended series of frames on the same project, both with and without the added 80% load on the CPUs? How much would that alter the GPU's PPD?

[toTOW]

It varies between project running on the GPU and the CPU applications involved ... in the worst case, on a 980 Ti, I'm loosing about 20k PPD.

A FAH CPU core is the worst case for the CPU (and 90% of the GPU projects are affected by a CPU load), so if my CPU slot won't produce more than 20k, I don't use the CPU.