DistrRTgen app_info file

[Maxwell]

I recently put in a new CUDA card, and put it on DistrRTgen to get to my 1M there. As I was futzing with it, I noticed how horrifically inefficient the app is. So, I spent a bit of time getting an app_info file running for the project - it is running the stock app, but has set the CPU priority to high and runs 2 WUs at once.

This is working on Win7x64, 6.10.58. Here it is, if anyone is interested:

<app_info>
 <app>
 <name>distrrtgen</name>
 </app>
 <file_info>
  <name>distrrtgen_3.20_windows_intelx86__cuda23.exe</name>
  <executable/>
 </file_info>
  <file_info>
  <name>cudart.dll</name>
  <executable/>
 </file_info>
  <file_info>
  <name>charset.txt</name>
 </file_info>

 <app_version>
  <app_name>distrrtgen</app_name>
  <version_num>320</version_num>
  <plan_class>cuda</plan_class>
  <avg_ncpus>0.75</avg_ncpus>
  <max_ncpus>1</max_ncpus>
 <flops>1.0e11</flops> 
  <coproc>
   <type>CUDA</type>
   <count>0.5</count>
   </coproc>
   <cmdline>p3 b-1</cmdline>
  <file_ref>
   <file_name>distrrtgen_3.20_windows_intelx86__cuda23.exe</file_name>
   <main_program/>
  </file_ref>
   <file_ref>
   <file_name>cudart.dll</file_name>
   <copy_file/>
  </file_ref>
   <file_ref>
   <file_name>charset.txt</file_name>
   <copy_file/>
  </file_ref>
  </app_version>
</app_info>

[zombie67]

How much does it improve performance?

[Maxwell]

How much does it improve performance?

I didn't get much of a baseline in terms of credits (pending, impatience, etc). But I'm watching the GPU load. Running one at a time, the typical load was about 50% - now it's >95%. Completion times have not noticeably increased, either. I'll post my credit total at the end of the day.

[zombie67]

I only have a handful of tasks validated so far, but this is what I found:

GTX 580:
o Normal: Run time averaged 70 seconds and pay averaged 51. That works out to about 63,330/day.
o With app_info: Run time averaged 133 seconds and pay averaged 61. That works out to about 78,865/day.

GTX 295:
o Normal: Run time averaged 352 seconds and pay averaged 1231. That works out to about 30,234/day.
o With app_info: Run time averaged 524 seconds and pay averaged 61. That works out to about 40.057/day.

The odd bit is around the credits awarded. Normally, they vary between 48 to 54 per task on the 580. But with the app_info, they are all awarded the exact same 60.74 credits. Similar situation with the 295, all exactly 60.74.

[Maxwell]

That's really interesting - I am not seeing the same behavior. I'm seeing huge variation in the number of credits awarded, from ~40 to ~142, with an average of 60.79 (damn close to your fixed value).

Since you did it, I went in and calculated numbers, and I'm seeing a similar gain. My number of normal WUs is fairly small, but I was getting 48k/day on a 570 normal, and 59k/day with the app_info file.

[zombie67]

Just curious, why did you pick .75 CPU per? With my GTX 295, that is taking 3 of my 4 CPUs...not leaving much for crunching. Can this be lowered at all, without impacting the performance? Anyone feel like doing some experiments?

[Maxwell]

I used 0.75 because the default was 0.85, and 0.75 seemed rounder - completely arbitrary, but I wanted to stick close.

For the record: At exactly 06:00 UTC 27 March 2011, I changed it over to 0.25. Anything reported after that time has the new app_info system...

[zombie67]

I used 0.75 because the default was 0.85, and 0.75 seemed rounder - completely arbitrary, but I wanted to stick close.

For the record: At exactly 06:00 UTC 27 March 2011, I changed it over to 0.25. Anything reported after that time has the new app_info system...

I think the "default" is based on the stats of the GPU. The faster the GPU, the higher the percentage of CPU required (to keep up, I assume). For example, the default of my 295 is .69, and for my 9500GT PCI is .27.

[Maxwell]

After dropping the CPU usage down to 0.25, I haven't seen a difference. I'm calculating around 58k/day now, well within the margin of error for the number of WUs I've run.

[Maxwell]

After dropping the CPU usage down to 0.25, I haven't seen a difference. I'm calculating around 58k/day now, well within the margin of error for the number of WUs I've run.

I think you're wrong, Maxwell...

After running it for a while at 0.25, I was noticing a severe lack of GPU loading on average - it was hovering around 55% on average. I changed it back up to 0.50 CPU/WU, and (running two WUs) this leaves 1 core dedicated to feeding the GPU. Since I made this change, my GPU load is above 95% on average. Yay!