Is it possible to use a specific GPU only for certain tasks?
23 May 2021 21:13:13 UTC
Topic 225456
(moderation:
Hi
I have a new Nvidea GTX1660S and a older Nvidea Strix GTX950 on my Win10 PC. At the moment I only use the 1660S. But on that GPU it takes a long time to finish WU's from Einstein@home while before on the GTX950, it was much quicker. Now I wonder if it is possible to configure Boinc in a way that Einstein only uses the GTX950 and other applications like Milkyway@home, uses the GTX1660S.
I only can find articles about running a application on 2 GPU's simultainusly or running multiple WU's on 1 (or more) GPU's but that's not what I'm looking for.
I'm looking forward to some responses.
Jos van der Vliet
karbonade wrote:HiI have
)
Yes you use an exclude part in your app_config file, something like this should work:
<cc_config>
<options>
<use_all_gpus>1</use_all_gpus>
\<exclude_gpu>
<url>http://einstein.phys.uwm.edu/</url>
<device_num>0</device_num>
</exclude_gpu>
\<exclude_gpu>
<url>http://milkyway.cs.rpi.edu/milkyway/</url>
<device_num>1</device_num>
\</exclude_gpu>
</options>
</cc_config>
That tells your system to NOT use gpu #0 from Einstein and exclude gpu #1 from Milkyway, change the settings to fit your system. To figure out which gpu is #0 and which is #1 by stopping and restarting Boinc and then going into the Event Log and seeing which gpu is listed as which number.
One problem I found over time is that you need to use the project names from the Add Project list in the Boinc Manager if you want to use any other Project.
Something sounds off. a
)
Something sounds off.
a 1660 Super should be MUCH faster than a 950.
_________________________________________________________________________
The best documentation to
)
The best documentation to help do the sort of things you mention is the BOINC User Manual. You can skip past basic stuff and find where the more advanced topics are explained.
There is a chapter on Client Configuration which goes through lots of options you can use in a configuration file (called cc_config.xml) you need to create. Some of these options will allow you to do the things you mention.
However, before you do that, there are things you mention that don't seem quite right. If you give more information, maybe your performance issues can be fixed.
Is that because you have removed the other one? If both are still installed, BOINC will use what it thinks is the 'best' of the two. The documentation will show you the option to <use_all_gpus>, which I guess you will need later if you want to run both for different projects.
This immediately raises a question mark. I don't use nvidia GPUs at at all but I would imagine that a 1660S should outperform a GTX950. I looked at your tasks list and saw GRP GPU tasks only, which were taking nearly half an hour. You don't mention if you are running concurrent GPU tasks so I assume you run them one at a time. For comparison, I have a very old (2013) AMD HD7850 that can complete 2 current GRP tasks in less than your time, so something seems not quite right.
If you do run tasks singly, perhaps your slow running GPU may be caused by too many CPU tasks running as well. Your CPU shows as 4C/8T. How many CPU tasks are you running? If it's "as many as possible", have you looked at what happens if you restrict that number a little - say run one or two less? You might get quite a speed-up in the crunch times for that GPU.
Look at the client configuration documentation for the explanation of the <exclude_gpu> option. It's a little more complicated than other options and you will need more than a single option - at least one for each different project that needs to use a particular GPU. There's a good example there showing how to specify both the project and the particular application (if necessary).
However before doing any of that, it would be useful to work out why your 1660S seems to be slow.
Cheers,
Gary.
To find the problem that
)
To find the problem that task
had, I would pull the GTX950 and let the GTX1660S try another WU.
If the problem persists, I would re-install the NVIDIA driver (since a Win10 update likes to "kill" GPU drivers), do a re-boot, and re-try a WU for the 1660S (the 950 still pulled).
If OK, then, if you have not yet done so, apply the "use_all_gpus" option, insert the 950 (in addition to the 1660S) and try running WUs.
If still no "success", try re-installing the NVIDIA driver again (newest one being 466.47), with both GPUs installed, and re-test.
Don't forget to put a check in the box at the bottom, when installing the driver, where it says "do a clean install".
I had a GTX960 running happily together with a GTX1650 - no problem ...
Thank you all for the quick
)
Thank you all for the quick replies.
Enough hints and tipt to go along with. I will reinstall the driver en make it a clean install. See what that does.
And I wil dig into the Client Information link as provide by Gary.
Thank you all.
Hi, Everything is now
)
Hi,
Everything is now working fine. The GTX1660S is almost 2X faster than the GTX950.
I performed a clean install of the drivers. That did do the trick.
Thank you all for your support!!!