My GPU is loaded at 95% or less in most wu. I would like to write an app info file (or should it be called an app config?) So that the gpu processes two wu simultaneously, but I don’t know how to do it. Maybe someone has a finished file?

My GPU is loaded at 95% or less in most wu. I would like to write an app info file (or should it be called an app config?) So that the gpu processes two wu simultaneously, but I don’t know how to do it. Maybe someone has a finished file?

App_config would be easier. Details are in the User Manual

Does that actually help, though? I haven't seen any obvious improvement, at least not on Windows.

Maybe someone has a finished file?

<app_config>
<app>
    <name>acemd3</name>
    <gpu_versions>
        <cpu_usage>1.0</cpu_usage>
        <gpu_usage>0.5</gpu_usage>
    </gpu_versions>
</app>
<app>
    <name>acemdlong</name>
    <gpu_versions>
        <cpu_usage>1.0</cpu_usage>
        <gpu_usage>0.5</gpu_usage>
    </gpu_versions>
</app>
<app>
    <name>acemdshort</name>
    <gpu_versions>
        <cpu_usage>1.0</cpu_usage>
        <gpu_usage>0.5</gpu_usage>
    </gpu_versions>
</app>
<app>
    <name>QC</name>
    <max_concurrent>8</max_concurrent>
</app>
<app_version>
    <app_name>QC</app_name>
    <plan_class>mt</plan_class>
    <avg_ncpus>4</avg_ncpus>
</app_version>
<app>
    <name>QC_beta</name>
    <max_concurrent>1</max_concurrent>
</app>
<app_version>
    <app_name>QC_beta</app_name>
    <plan_class>mt</plan_class>
    <avg_ncpus>4</avg_ncpus>
</app_version>
<project_max_concurrent>4</project_max_concurrent>
</app_config>

I would be surprised if it resulted in higher performance.

My gut says at most it might increase 5% from filling small gaps.
I was thinking of testing it but the problem is the RAC is a 10-day average. Therefore I need to run without changes for over 10 days to establish a baseline. Double up and then run over 10 days to see if it changed. And, hope that you didn't increase your protein length in the mean time.

All systems have different sizes and therefore efficiencies already.

What's a "system" to you?
You know we're all very curious what you're up to.
I opened one of your files trying to double click app_config and saw a very long epitope or a very short protein. The amazing thing was it had a dozen prolines in a row. I've been trying to imagine what conformation that would have and concluded a pig tail.

My gut says at most it might increase 5% from filling small gaps.
I was thinking of testing it but the problem is the RAC is a 10-day average. Therefore I need to run without changes for over 10 days to establish a baseline. Double up and then run over 10 days to see if it changed. And, hope that you didn't increase your protein length in the mean time.

I ran that experiment, a while back when the new app came out, and running 2 WUs at once reduced productivity. See link:


http://www.gpugrid.net/forum_thread.php?id=4989&nowrap=true#52749


You don't need 10 days to see the results. Just run 2 WUs simultaneously for a few minutes on one machine. In Boinc Manager, under task bar, highlight the tasks and click properties, and see the progress rate. Then you can suspend one of the tasks, and see the change in the progress rate, in the task still running after a few minutes. You can also compare that to your base line of running one unit at a time. It is not an exact science, especially since the unit vary in length. but it will give you a good idea of what the difference is.

Your test is too short. The biggest gain when running 2 WUs per GPU is the gap when a WU finishes. I think one needs to test for a few WUs.Be nice if one used the same pair of WUs.

I thought of an easier way to test it. The Statistics graph in BOINC Manager. I'll run for a few days with singletons and then double up and see if the slope changes.

Your test is too short. The biggest gain when running 2 WUs per GPU is the gap when a WU finishes. I think one needs to test for a few WUs.Be nice if one used the same pair of WUs.

I thought of an easier way to test it. The Statistics graph in BOINC Manager. I'll run for a few days with singletons and then double up and see if the slope changes.

Good, test it. Maybe results for this batch will be different than the previous batch. Let's know what you find.

There's a problem that prevents a good test. With only 2 WUs per GPU I will constantly have several GPUs running one or none.

If Toni sees fit to increase the limit to 3 WUs per GPU a test would work.