I have been running a GFN-WR task for over a day now, but the CPU usage has not dropped below 97%. I am using Gentoo Linux. I installed the closed-source Nvidia drivers instead of the open source ones. Is there any reason for the high CPU usage?

I have been running a GFN-WR task for over a day now, but the CPU usage has not dropped below 97%. I am using Gentoo Linux. I installed the closed-source Nvidia drivers instead of the open source ones. Is there any reason for the high CPU usage?

Do you mean telling Xserver to run on integrated graphics? I'm not even running an Xserver right now and there's no display connected to the computer.

Do you mean telling Xserver to run on integrated graphics? I'm not even running an Xserver right now and there's no display connected to the computer.

I think that fixed it, though my way is really hacky. I ended up having to export the environment variable mentioned in the post you linked before starting boinc, meaning everything boinc runs will be started with that variable.

Now, is there a way I can tell boinc to only start the gfn applications like that? I know about command line options in an app_config file, but does that work with environment variables?

I think that fixed it, though my way is really hacky. I ended up having to export the environment variable mentioned in the post you linked before starting boinc, meaning everything boinc runs will be started with that variable.

Now, is there a way I can tell boinc to only start the gfn applications like that? I know about command line options in an app_config file, but does that work with environment variables?

Do you mean telling Xserver to run on integrated graphics? I'm not even running an Xserver right now and there's no display connected to the computer.

Yes but if there's no display connected this can't resolve the problem.
The CPU usage is close to zero with the same code running on Windows... why?
I found https://bitcointalk.org/index.php?topic=181328.0
but I don't have access to a Linux box for testing.

I can test this on my system, I've been messing with GFN code and I have the Makefile set up the way I want. What branch should I checkout?

I think that fixed it, though my way is really hacky. I ended up having to export the environment variable mentioned in the post you linked before starting boinc, meaning everything boinc runs will be started with that variable.

Now, is there a way I can tell boinc to only start the gfn applications like that? I know about command line options in an app_config file, but does that work with environment variables?

You don't need to use the YIELD_SLEEP_TIME environment variable. The default value is hardcoded, so you can change that when you compile the preload library. The lib_init function is just there for flexibility, so you can omit the lib_init function if you want.

I think that fixed it, though my way is really hacky. I ended up having to export the environment variable mentioned in the post you linked before starting boinc, meaning everything boinc runs will be started with that variable.

Now, is there a way I can tell boinc to only start the gfn applications like that? I know about command line options in an app_config file, but does that work with environment variables?

You don't need to use the YIELD_SLEEP_TIME environment variable. The default value is hardcoded, so you can change that when you compile the preload library. The lib_init function is just there for flexibility, so you can omit the lib_init function if you want.

I still have to have the LD_PRELOAD variable though, unless there's another way of doing it I don't know.

On Linux, it doesn't look good for small GFN, but there is tunable potential to drive the GPU to 100% with low CPU utilization for large GFN.

Testing was with a 3.3 GHz 5820K with a GTX 760. In all cases, without a microsecond sleep, a GFN task takes 100% of a CPU thread.

Then I added a microsecond sleep in HostProgram.h after each call to clEnqueueNDRangeKernel(). So:
- one right after that function call in Execute(), and
- another after the error check following the function call in ExecuteProfiling().

I ran tests with 272042236^(2^N)+1 and this selected OCL2 in every case. These are the estimated runtimes reported by the program.

The CPU has a hard time feeding the GPU when GFN workunits are small, so introducing any sleep for GFN17, or smaller, lengthens the GPU computation time while reducing GPU and CPU utilization.

Microsecond sleeps cause interesting things to happen to the estimated runtime and the % of CPU and GPU for GFN18 and above. With a 1 microsecond sleep, we start to see significant reduction in %CPU utilization.

Thanks for that data. So, from your results, the default value of YIELD_SLEEP_TIME=1000 seems to be pretty good? And the larger the workunit, the higher value YIELD_SLEEP_TIME should be? On a side note, I'm now running AP27 tasks on my gpu with the same library, and it's only using ~2.9% cpu. I don't have data to compare that with right now, but it seems like that would be ideal.

Your best value of YIELD_SLEEP would be different than mine. It depends on your hardware and the application you run. In my case 1000 is a not a good value for any GFN task. It's way too much for GFNs smaller than GFN22, and not enough for GFN22 or higher. My "best" value for GFN22 is 1610 (all these are found by trial and error: initial guess then bisection). GFN could be altered to dynamically find it's own best values, doing so with several iterations of tests, by setting a target runtime of (for example) 0.5% over the no-sleep runtime.

Okay, one last question: BOINC shows CPU usage for each task, but how did you figure out the GPU usage?

Okay, one last question: BOINC shows CPU usage for each task, but how did you figure out the GPU usage?

And, if it is running as a service, edit /lib/systemd/system/boinc-client.service and modify the following line
ExecStart=/bin/sh -c 'YIELD_SLEEP_TIME="500" LD_PRELOAD="/your/libsleep/path/libsleep.so" /usr/bin/boinc --dir /var/lib/boinc-client >/var/log/boinc.log 2>/var/log/boincerr.log'

CPU reduced from 100% to less than 20%.