POEM@home News

[mmstick]

You pretty much have it, for the 4870 I simply have the avg_ncpu to 0.67, max_ncpu to 3, and count to 0.33. Using such gives 100% utilization. 6850 though I can't seem to get 100% CPU, but can at least get 92% using (now) either the 10 WUs at once or this:

<app_info>

<app>
<name>poemcl</name>
<user_friendly_name>POEM++ OpenCL</user_friendly_name>
</app>

<file_info>
<name>poemcl_0.1_windows_intelx86__opencl_ati_10 0</name>
<executable/>
</file_info>

<app_version>
<app_name>poemcl</app_name>
<version_num>1</version_num>
<plan_class>opencl_ati_100</plan_class>
<avg_ncpus>0.67</avg_ncpus>
<max_ncpus>4</max_ncpus>
<flops>2.1e10</flops>
<coproc>
<type>ATI</type>
<count>0.16</count>
</coproc>
<cmdline></cmdline>
<file_ref>
<file_name>poemcl_0.1_windows_intelx86__opencl_ati _100</file_name>
<main_program/>
</file_ref>
</app_version>

<app>
<name>poempp</name>
<user_friendly_name>POEM++</user_friendly_name>
</app>

<file_info>
<name>poempp_0.8_windows_intelx86</name>
<executable/>
</file_info>

<app_version>
<app_name>poempp</app_name>
<version_num>8</version_num>

<file_ref>
<file_name>poempp_0.8_windows_intelx86</file_name>
<main_program/>
</file_ref>
</app_version>
</app_info>

The extra app I added to it recently is so that I can do both CPU and GPU WUs at the same time, so that I may run 4 cores in CPU, with the other 4 cores running my 6850.

I might buy two 7970's or the dual version in next few months, I would like to see how well they perform, and hopefully a price drop soon. If they perform as well as early GPGPU benchmarks have shown, they are going to be awesome in GPU projects, once they are supported.

[Dirk Broer]

As I am a noob regarding app_info.xml files, I have some questions when I compare mine with yours:

<avg_ncpus>0.67</avg_ncpus>

Meaning that on average .67 of your 8 cores is to be used? that would be 5.36 core so that can not be right.
.67 of 4 cores (see below) is 2.68, so I am missing someting here.
I think my own 0.5 means that 1 core of that cpu is to be used, and it does.

<max_ncpus>4</max_ncpus>

Here you max the total of cores to be used at 4. As you have an octo-core you can do that.
My 0.5 would not work then, but it does, so I am missing someting here.

<flops>2.1e10</flops>

I've also seen app_info.xml stating 4.1e12, would that speed up the processing, or will the WUs be bigger?

<count>0.16</count>

Meaning 6 POEM WUs at a time on the one core of your GPU? 1/6=roughly 0.16, but better use 0.1666666666667 then
My 0.5 means 2 Wus at a time, and it does.

[mmstick]

0.67 simply means it is using 0.67 of 1 CPU core per unit on average. So avg_ncpus(0.67) * numberOfGPUWUs(6) = 4 total cores used.

For max_ncpu I have never seen it affect anything at all, I suppose it was meant to able to speed up a WU to the number set if the processor is otherwise idle and capable of going beyond average.

For the FLOPS I simply have no idea. Overall there seems to be a lack of documentation on how to properly configure and what each parameter means in BOINCs app_info.xml.

And yes, the 0.16 is the number to use for 6 WUs, since the program doesn't understand anything beyond two decimal points and if you set 0.17 it seems to refuse to run more than 5.

[Dirk Broer]

So when I run 2 WUs on my HD 4770 as I do now, the
<avg_ncpu>0.50</avg_ncpu> means 50% of a cpu core is available to support GPU crunching (per WU I hope),
<max_ncpu>0.50</max_ncpu> seems pretty pointless while doing two WUs at a time given this argument from the site of the link before:
"<max_ncpus></max_ncpus> - This controls the total amount of CPU core(s) available to support GPU crunching. Usually for best performance it should represent the number of concurrent tasks specified in <count> value multiplied by the amount of CPU core available to support each task as specified in <avg_ncpus> value. Therefore for a single concurrent task <avg_ncpus> value and <max_ncpus> value is the same but <max_ncpus> value is a multiple when more than 1 task is run concurrently"
Mine should be 1 then.

<flops>2.1e10</flops> means the expected FLOPS of the application

<count>0.50</count> means that 2 WUs will be proccessed at a time per GPU core
<count>0.33</count> would mean three tasks

[Nolion]

Can anyone share an example of a functioning Linux x86_64 app_info.xml? I'm getting error messages regarding the application filename when I try to configure for POEM++ OpenCL version v0.02 (opencl_ati_100). This is with Ubuntu 11.04, BOINC 7.0.2, and HD4850. Grateful for any help.

[vaughan]

Welcome Nolion. Not sure if this is what you need but it helped me in Windows 7.

From message 5323 at the official Poem forum:
Go here: http://boinc.fzk.de/poem/download at the bottom...
Sometimes takes a while to load in the browser, so be patient...

[Dirk Broer]

Hi Nolion,

Try this and post the result:

<app_info>
   <app>
      <name>poemcl</name>
      <user_friendly_name>POEM++ OpenCL</user_friendly_name>
   </app>

   <file_info>
      <name>poemcl_0.02_x86_64-pc-linux-gnu</name>
      <executable/>
   </file_info>
 
   <app_version>
      <app_name>poemcl</app_name>
      <version_num>1</version_num>
      <plan_class>opencl_ati_100</plan_class>
      <avg_ncpus>0.5</avg_ncpus>
      <max_ncpus>1.0</max_ncpus>
      <flops>2.1e10</flops>
      <coproc>
         <type>ATI</type>
         <count>0.5</count>
      </coproc>
      <cmdline></cmdline>
      <file_ref>
         <file_name>poemcl_0.02_x86_64-pc-linux-gnu</file_name>
         <main_program/>
      </file_ref>
   </app_version>

</app_info>

[Nolion]

Hi Nolion,

Try this and post the result:

<app_info>
   <app>
      <name>poemcl</name>
      <user_friendly_name>POEM++ OpenCL</user_friendly_name>
   </app>

   <file_info>
      <name>poemcl_0.02_x86_64-pc-linux-gnu</name>
      <executable/>
   </file_info>
 
   <app_version>
      <app_name>poemcl</app_name>
      <version_num>1</version_num>
      <plan_class>opencl_ati_100</plan_class>
      <avg_ncpus>0.5</avg_ncpus>
      <max_ncpus>1.0</max_ncpus>
      <flops>2.1e10</flops>
      <coproc>
         <type>ATI</type>
         <count>0.5</count>
      </coproc>
      <cmdline></cmdline>
      <file_ref>
         <file_name>poemcl_0.02_x86_64-pc-linux-gnu</file_name>
         <main_program/>
      </file_ref>
   </app_version>

</app_info>

Hi Dirk - Thanks for the suggestion even though it didn't go well for me. This is what I got in my event log (you're welcome to point out any possible newbie mistakes I made that may be contributing):

Mon 23 Jan 2012 08:22:51 PM EST | http://boinc.fzk.de/poem/ | Found app_info.xml; using anonymous platform
Mon 23 Jan 2012 08:22:51 PM EST | http://boinc.fzk.de/poem/ | File referenced in app_info.xml does not exist: poemcl_0.02_x86_64-pc-linux-gnu
Mon 23 Jan 2012 08:22:51 PM EST | http://boinc.fzk.de/poem/ | [error] State file error: missing application file poemcl_0.02_x86_64-pc-linux-gnu
Mon 23 Jan 2012 08:22:51 PM EST | Poem@Home | [error] No application found for task: x86_64-pc-linux-gnu 2 opencl_ati_100; discarding

[Nolion]

I got past the error messages thanks to suggestion from vaughan (download file from poem) combined with Dirk's app_info.xml (edited filename to match the downloaded file). Now just waiting to see if the project will send some tasks. This is the app_info.xml currently running...

<app_info>

   <app>

      <name>poemcl</name>

      <user_friendly_name>POEM++ OpenCL</user_friendly_name>

   </app>



   <file_info>

      <name>poemcl_0.2_x86_64-pc-linux-gnu__opencl_ati_100</name>

      <executable/>

   </file_info>

 

   <app_version>

      <app_name>poemcl</app_name>

      <version_num>1</version_num>

      <plan_class>opencl_ati_100</plan_class>

      <avg_ncpus>0.5</avg_ncpus>

      <max_ncpus>1.0</max_ncpus>

      <flops>2.1e10</flops>

      <coproc>

         <type>ATI</type>

         <count>0.5</count>

      </coproc>

      <cmdline></cmdline>

      <file_ref>

         <file_name>poemcl_0.2_x86_64-pc-linux-gnu__opencl_ati_100</file_name>

         <main_program/>

      </file_ref>

   </app_version>



</app_info>

[mmstick]

You'll find the performance to be rather terrible in Linux as compared to Windows, especially if you want to run any CPU tasks alongside the GPU one since for some reason any CPU task will starve the POEM OpenCL work units of CPU time, even if most of your processor is being spent idle. Example, a configuration I used in Windows got 90% GPU usage along with four World Community Grid tasks at the same time. In Ubuntu, I received 70% GPU usage with the same app info without CPU tasks, and 50% usage as soon as it started doing World Community Grid tasks, despite having 60% of my entire CPU being spent idle. I have no idea why this happens in Linux, maybe it will be fixed.