I world about the new GPU tasks.
10 Mar 2023 6:27:01 UTC
Topic 229204
(moderation:
I think the points awarded for the "Multi-Directional Gravitational Wave search on O3 (GPU) v1.00 () x86_64-pc-linux-gnu" tasks are bit crap considering how they lock the the GPU from new tasks while the CPU has process the rest of the data.
My Quadro RTX A4500 spends about 2 minutes computing and then is forced to be idle for 3.3 minutes while the CPU does it's part.
I don't mind donating my hardware and electricity bill but the fact that I can't reach optimal performance because of the code doesn't mean I should I penalized. The task of CPU cores in this use case could (in theory) be done by copying the data out of VRAM and letting a separate CPU core process it while the GPU moves on to the next task. Barring that, I suggest increasing the points rewarded for the these work units.
Run several of these tasks
)
Run several of these tasks concurrently. Your gpu will "always" have something to work on.
cuphi wrote: I think the
)
Boca Raton gave you the answer but to add to that I pause the 2nd task for a little bit so they aren't both starting and stopping at the same time, after you get it synced like that the gpu should just keep on crunching.
I'm getting the same GPU
)
I'm getting the same GPU under-utilization on my MacPro5,1 with a Radeon 570 and 2 Radeon 560s (a bit of kit-bashing to get all 3 to fit). My daily credits have taken a nose dive. When I look at the GPU utilization there are long gaps near 0% and very few 100% times. Prior to these tasks all 3 GPUs were near 100% constantly.
Do I run these tasks concurrently by
1) changing the project pref "GPU utilization factor of GW apps: " to 0.5, or
2) by creating this app_config:
<app_config>
<app>
<name>einstein_O3MDF_1.00_x86_64-apple-darwin__GW-opencl-ati</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>
</app_config>
TIA
Either/Or works. Less work to
)
Either/Or works. Less work to change the utilization for gpu apps in the project preferences.
Though I believe the website preference change route won't give a full cpu thread for support.
Usually somewhere around 0.9 cpu threads. Only a factor in scheduler resource allocation. Won't impact the gpu application in any way.
CyborgSam wrote: I'm getting
)
The 03 gpu tasks only give 1000 credits per task, the other gpu tasks pay @3333 credits per task, so if you were running them instead of the 03 gpu tasks it makes sense your RAC will nosedive. Mine sure did when I started running these and yes mine have the same LONG gap where they are doing nothing, they really are but it's not on the gpu, and then it picks back up again with the next task starts.
I went to Project Preferences
)
I went to Project Preferences and unchecked both Multi-Directional Gravitational Wave search on O3 (GPU) and Multi-Directional Gravitational Wave search on O3 (CPU), saved my preferences, and updated the project in BOINC. It took over a day to clear out all the queued tasks. Things went great for about a day...
Today I'm running the GPU tasks again and have another day or two of them queued up. I checked my project preferences, they're correct. I didn't restart or quit/relaunch BOINC.
WTF?
I wouldn't mind doing these tasks if I got decent credit, but my Mac Pro's RAC is about 1/3 or less of what it was before it started these tasks.
How do I make them stop for good?
OK, this might be my bad. I neglected to remove the app_config.xml file from Einstein after changing the project prefs. I just removed it, I'll have to wait about a day for the GW tasks to finish.
Am I correct in assuming that the local app_config.xml overrode the project settings?
CyborgSam wrote:OK, this
)
It's complicated.
Important details about the app_config.xml mechanism are covered in the documentation. In particular, note where it mentions, "If you remove app_config.xml, or one of its entries, you must reset the project in order to restore the proper values."
This is because the contents of that file are inserted into the state file (client_state.xml). Removing the app_config.xml file doesn't remove what was inserted. If you understand the structure of the state file you could manually edit it out but the project reset is the safe way to get it removed.
Another potential 'gotcha' is the preference setting for allowing 'Non-preferred apps'. Make sure that is set to 'No'. Otherwise (if set to 'Yes') the scheduler might decide to send tasks for a 'non-preferred app' basically on a whim :-).
Cheers,
Gary.
Gary-> Thanks. I changed
)
Gary->
Thanks. I changed the project pref for Non-preferred apps to No. In BOINC I set Einstein to No new tasks so I can safely reset the project after they all complete. The Mac Pro does work for other projects so it should keep itself 100% busy until I reset Einstein (it'll probably happen in the middle of the night).
I looked in client_state.xml, I found part of what was in app_config but didn't quickly see all of it. So I'll play it safe.
Wanted to 2nd the motion on
)
Wanted to 2nd the motion on increasing the points for these particular tasks.
My computer currently has no preferences over what tasks it pulls. However I did notice the "Multi-Directional Gravitational Wave search on O3 (GPU) 1.03 (GW-opencl-nvidia)" is quite the heavy hitter on the GPU resources. Seems to require many times the amount of vram as other tasks (about 5.5gb per task according to my BOINC client). I have a 12gb 3080TI and 3 of these tasks running at the same time will consume about 95% of all available vram. The run time on these tasks vs. other GPU tasks is about the same. However other tasks pay out 3k points and these pay out 1k in points. Seems a little unoptimized and from my perspective. I would think if I'm optimizing for PPD. I would want to exclude these tasks from my PC.