Gravitational Wave Search O3 tasks fail on ATI iGPU, work on nVidia GPU
28 Jul 2021 5:25:44 UTC
Topic 225739
(moderation:
I have a laptop with an ATI iGPU and an nVidia discrete GPU, which has been getting Gravitational Wave Search tasks. It seems that all or nearly all of the tasks for the ATI iGPU fail like this one: https://einsteinathome.org/task/1149663510 while the tasks for the nVidia GPU generally succeed.
I'd like to either fix the problem that causes the tasks to fail on the ATI iGPU, or if that's not really doable, find some way so that the computer doesn't get ATI O3 tasks, and gets other tasks for the ATI iGPU, while continuing to get any available tasks for the nVidia GPU.
alee67 wrote: I have a
)
You can unselect which gpu you do not want to get tasks for:
I only want to exclude the
)
I only want to exclude the Gravitational Wave Search O3 tasks for ATI, not for nVidia, and not other tasks for ATI. Is it possible to do this, perhaps in app_config.xml? If I get the right app name, I can set <max_concurrent>0</max_concurrent>, but that may simply prevent the tasks from being run, while I'd like to prevent them from being retrieved altogether.
you "can" set it up for
)
you "can" set it up for certain devices to run certain apps exclusively, using the exclude options for the cc_config file.
however, the larger problem will be work fetch. it's very likely that you will eventually run into an issue where your cache will eventually be filled with only one type of work (most likely GW due to the DCF differences between these tasks) and your AMD integrated device will end up sitting idle more often than not.
the only real solution to this, is to run two separate BOINC client instances, to totally separate the work fetch and caches between the apps/devices. One BOINC client for Nvidia/GW, and another for AMD/GR. but this comes with the caveat of more complexity and split stats for a single host.
_________________________________________________________________________
Unless there's some way to
)
Unless there's some way to prevent all or nearly all GW tasks from failing with a computation error on the ATI iGPU, there's little point in getting them. Why would I need a complicated setup with two BOINC clients? The GPU tasks are specific to the type of GPU, whether Intel, ATI or nVidia.
I added the following to cc_config.xml:
<exclude_gpu>
<url>http://einstein.phys.uwm.edu/</url>
<type>ATI</type>
<app>einstein_O3AS</app>
</exclude_gpu>
I think I got the correct URL and app name to use from client_state.xml for this.
You’ll find out. When
)
You’ll find out.
When you have both apps enabled, the project tends to skew towards GW. And you’ll end up with a cache full of GW tasks (running on your nvidia card) and your AMD iGPU sitting idle (because it’s full of nvidia GW that you’ve excluded). Work fetch will end up reporting that your cache is full because there’s no way to set separate cache levels for different apps. Increasing cache size just moves the goal post until the process happens again.
many of us have tried this before.
_________________________________________________________________________
alee67 wrote:Unless there's
)
Yes, that's correct. Well done for solving your problem.
The GW GPU app couldn't be run without error on internal GPUs in the past so there's no reason to expect anything different with this latest app. Since you only have GW GPU tasks showing on the website, what you have done will be the best way to go. If you were to start adding the gamma-ray pulsar tasks (which may very well run on your internal GPU, albeit rather poorly) then you may have the sorts of problems mentioned. Since there's no evidence you're intending to do that, you should be fine now that you've excluded the badly behaving GPU.
Looks OK to me - not that I've ever needed to exclude a GPU before :-).
Cheers,
Gary.