Advanced search

Message boards : Graphics cards (GPUs) : I can run two tasks but not three

Author Message
WPrion
Send message
Joined: 30 Apr 13
Posts: 69
Credit: 1,000,072,969
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 42246 - Posted: 1 Dec 2015 | 1:31:19 UTC

Should I be able to run 3 (or more) tasks on the GPU?

I've been running one or two tasks on a 980ti with no problems. The card is not overclocked.
When I run one task the GPU load hovers around 66%, and with two tasks the GPU load hovers around 81%.

I'd like to try 3 tasks, but when I do I get this event log item:
"| GPUGRID | This computer has reached a limit on tasks in progress"
and no third task is downloaded; the two in progress keep running.

Here's my app_config:

<app_config>
<app>
<name>acemdlong</name>
<max_concurrent>3</max_concurrent>
<gpu_versions>
<gpu_usage>0.33333</gpu_usage>
<cpu_usage>0.33333</cpu_usage>
</gpu_versions>
</app>

Thanks,
Win

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2076
Credit: 15,122,124,983
RAC: 4,726,061
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 42247 - Posted: 1 Dec 2015 | 1:57:33 UTC - in response to Message 42246.
Last modified: 1 Dec 2015 | 2:01:33 UTC

Should I be able to run 3 (or more) tasks on the GPU?
Theoretically you can run 3 tasks simultaneously on the GPU, but there's a limit of 2 WU per GPU for a host here in GPUGrid.

I've been running one or two tasks on a 980ti with no problems. The card is not overclocked.
When I run one task the GPU load hovers around 66%, and with two tasks the GPU load hovers around 81%.
You should try to set the SWAN_SYNC environmental variable to reserve a CPU thread to feed the GPU a bit faster.
It's method is described in this post.

klepel
Send message
Joined: 23 Dec 09
Posts: 171
Credit: 2,852,040,338
RAC: 615,953
Level
Phe
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 42248 - Posted: 1 Dec 2015 | 2:10:04 UTC

I should not write it: Or you set up another project for the 19% you have spare.
Something like:
For GPUGRID.net:
<app_config>
<app> <name>acemdlong</name>
<gpu_versions>
<gpu_usage>0.4</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
<app>
<name>acemdshort</name>
<gpu_versions>
<gpu_usage>0.4</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
</app_config>

And as an example primegrid.com:
<app_config>
<app>
<name>genefer22</name>
<gpu_versions>
<gpu_usage>0.2</gpu_usage>
<cpu_usage>0.25</cpu_usage>
</gpu_versions>
</app>
<app>
</app_config>

And please do not forget to adjust your ressurse share to:
GPUGRID 100%
PRIMEGRID: 0%

You might have to adjust your config’s to get best results. Works for me when GPUGRID does up-load on of the WUs and I am not able to get an other one.

WPrion
Send message
Joined: 30 Apr 13
Posts: 69
Credit: 1,000,072,969
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 42249 - Posted: 1 Dec 2015 | 2:34:52 UTC - in response to Message 42247.

Well, that takes care of that!

I'll check out the SWAN_SYNC.

Thanks,

Win

WPrion
Send message
Joined: 30 Apr 13
Posts: 69
Credit: 1,000,072,969
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 42250 - Posted: 1 Dec 2015 | 3:16:43 UTC - in response to Message 42247.

Very interesting. Before I added the swan_sync I rarely had more than 8% cpu for each of the two tasks running on the GPU. (My cpu is 2 core/4 thread.)

Now with swan_sync, cpu is 25% for each task - one thread each. GPU load, however, has only gone from about 82% to 85%. If I run no other cpu task (from other projects) the cpu is still 25% for each GRID task, but the GPU load goes up to 90% which is the most I've seen so far.

Is all this additional cpu power going to support the gpu, or is the cpu doing it's own number-crunching on the grid tasks?

I briefly considered swapping the cpu for a 4core/8thread but didn't think it was worth it. Maybe I should reconsider.

If swan_sync is set to 0 does that essentially disable the variable (equivalent to deleting it) or does it have some other effect?

I will be looking for some way to throttle total GPU load. Right now in the cold season my "computer room" is very cool (GPU temp is only 32C). But this room heats up dramatically in the summer and I want to keep the GPU temp under control (I don't like getting close to any limits.) I was planning to simply go back to single task on the gpu - which brings it down to 65% load.

Are there more graceful ways to throttle total GPU load for temperature control?

Thanks,

Win

Vagelis Giannadakis
Send message
Joined: 5 May 13
Posts: 187
Credit: 349,254,454
RAC: 0
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwat
Message 42257 - Posted: 1 Dec 2015 | 9:50:00 UTC - in response to Message 42250.

The most effective way of controlling the temperature that I've found is downclocking. I reduce the clock at 100MHz steps, works beautifully!

Adjusting the power target is another means preferred by some, but I dislike it, as it tends to make the GPU frequency jump around all the time, which had caused some WUs to fail on me. It could however achieve the double goal of both reducing the temperature and maximizing efficiency, giving you some more performance and credits.
____________

WPrion
Send message
Joined: 30 Apr 13
Posts: 69
Credit: 1,000,072,969
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 42258 - Posted: 1 Dec 2015 | 13:07:44 UTC - in response to Message 42257.

Do you use Afterburner for the downclocking? I take it you type in negative numbers - I didn't think you can drag the sliders negatively.

Thanks,

Win

Post to thread

Message boards : Graphics cards (GPUs) : I can run two tasks but not three