The O2-All Sky Gravitational Wave Search on GPUs - discussion thread.

I had to abort all the GW GPU tasks on my LINUX/AMD machine. 2 more invalid results along with a high CPU warning on 2 other tasks. The NVIDIA/WIN 7 machine is handling the GW GPU task just fine, so far.

That seems to be the case across the board. Looking at my teammates, I noticed a RAC drop for everyone (even Gary).

Craig, are you running any CPU tasks as well, or, did you limit the box to GPU only? I am referring to the box where you have 1 GPU, running x3.

Dang, sorry that you are having to deal with that frustration. Yes, let's hope more information from other users helps solve the problem.

With my app_config file set up with this:

  <app>
     <name>hsgamma_FGRPB1G</name>
     <fraction_done_exact/>
     <gpu_versions>
       <gpu_usage>0.333</gpu_usage>
       <cpu_usage>0.25</cpu_usage>
     </gpu_versions>
   </app>
   <app>
     <name>einstein_O2AS20-500</name>
     <gpu_versions>
       <gpu_usage>0.333</gpu_usage>
       <cpu_usage>0.25</cpu_usage>
     </gpu_versions>
   </app>
  <project_max_concurrent>6</project_max_concurrent>

... my cc_config file set with these options:

    <ncpus>6</ncpus>                        #I bumped that up from the default 4.
    <use_all_gpus>1</use_all_gpus>
     <exclude_gpu>
         <url>einstein.phys.uwm.edu</url>>
         <device_num>1</device_num>
         <app>einstein_O2AS20-500</app>
     </exclude_gpu>
    <exclude_gpu>
        <url>einstein.phys.uwm.edu</url>
        <device_num>0</device_num>
        <app>hsgamma_FGRPB1G</app>
    </exclude_gpu>

...and Project preferences set accordingly to run both apps, I thought I would be able to run both FGRPB1G and GW GPU tasks simultaneously, each partitioned to it's own GPU (both RX 570s). Alas, no.

What happened was that only one GPU ran tasks. My queue started out with only GW GPU tasks. When those completed (nothing else was downloading), I reset Project prefs to run just the FGRPB1G app. Then that same GPU ran only FGRPB1G tasks. The inactive GPU was recognized by the system and two GPUs (RX570s) were listed on the host's E@H computer page. The unused GPU was 'alive', just not getting work. So I removed the exclude_gpu option, restarted boinc, and through a hazy series of steps eventually got both GPUs running tasks again. Interestingly (or perhaps not), when Project prefs are set to run both apps, only GW GPU work is downloaded.

As far as running both FGRPB1G and GW GPU on separate GPUs, was I trying to do the impossible, the inadvisable, or the possible-but-missing-a-step?

For some days I have suspected that there is something inconsistent or time-varying in the actual display of "inconclusive" task status at the Einstein account web pages.  I have more than once spotted one or two inconclusives, only to find none a few minutes later.  At the time I made the quoted post, my pending task list for my first host showed well over a dozen (the cited 11 run at 3X, plus many more run at 4X).  And yet, most times since, I see none.

Sadly, this makes checking one's pending list for inconclusive status tasks as an early warning of yet-to-come invalid result findings an iffy business.  If you see them, you probably have a problem.  If you don't see any, this may just not be a moment when the site software puts them on view for you.

And, yes, with another day gone by the finding that for my Windows 10 RX 570 hosts running 1.07 GW at 2X works fine, but 3X and 4X give somewhere between extremely high and 100% invalid rates continues to be borne out by a rapidly rising pile of data.