HELP - Why all my GPU tasks fail
23 Oct 2021 19:07:39 UTC
Topic 226286
(moderation:
So,
Very frustrated...
I have multiple computers, all with petty old NVIDIA GPUs (GTX 660 TIs, 770s, etc.) that have 2 GB memory or smaller. Most if not every single GPU task fails. I think once and a while one does complete without a computation error.
Obviously right now I cannot buy new GPUs due to insane prices and availability (basically non-existent).
I'm trying to figure out and alleviate/solve whatever is wrong. I think previously someone mentioned einstein@home GPU tasks now require more memory than 2 GB?
If that's case, why would BOINC keep giving me tasks when it knows I cannot do them? (I don't know if this is 100% true or not)
It's driving me crazy.
Today am I extra frustrated because my BOINC manager refuses to take me to ay sites when I click on say "you messages" or "your results" buttons, etc. On one of my machines (where I am writing this)
I've tried numerous ways to just stop using the GPU, to no avail.
I keep getting tons of tasks, failing them all and wasting both my time and resources, and the Einstein servers. Also my machine max out on jobs for the day because of all the downloads/failures, I think it's also negatively affecting my credit.
Anyone care to help me determine what need to systematically do or atl east give guidance how to turn off my GPU tasks?
I've tried numerous ways through the GUI and app_config.xml, currently I have a cc_config.xml with the following that's not stopping GPU tasks.
<use_all_gpus>0</use_all_gpus>
The issue is the GW tasks
)
The issue is the GW tasks with your low GPU memory. Einstein has two different kinds of tasks; Gravitational Wave search, and Gamma Ray Pulsar Binary Search. You need to go into your preferences and disable the processing of the Gravitational Wave O3AS tasks. These tasks almost always need more than 2GB of GPU memory.
you will only really be able to process the Gamma Ray tasks on GPUs with such low memory.
disable GW O3 All-Sky and your issues will go away.
change them here: https://einsteinathome.org/account/prefs/project
scroll down and uncheck the GW O3 All-sky button.
<use_all_gpus>0</use_all_gpus>; this setting doesn’t do what you think it does. This setting controls whether BOINC will use all GPUs when multiple types/models are installed. With this set to 0, BOINC only will use the best GPU in the system (or all if they are all the same). If you have different models installed, and you want BOINC to use all of them, you need to leave this on ‘1’
_________________________________________________________________________
Thanks much for this. About
)
Thanks much for this. About to try what you said, though I am very confused why that would work.
So I just want to disable the single app Gravitational Wave search O3 All-Sky? Not sure how that's affecting the graphics tasks...
Both my machines 11/16/2021 (before making any changes, with all apps are enabled) currently only have work for:
GW search O2 Multi-directional 2.09 (GWNew)
Gamma-Ray Pulse Search #5 1.08 (FGRPSSE)
and nothing else.
After I disabled the above app, and re-enabled the GPU and did an update (with gpu util factors all at 1.0) I now am getting tasks: Gamma-ray pulsar binary search #1 on GPUs 1.22 (FGRPopencl-nvidia)
I see they are running a single one of those tasks at a time (good).
We'll see if they work without failing. It's possible in the past since I had the factors at 0.5 that there wasn't enough memory and they were failing, but I'm pretty sure I had set them back to 1 and were still failing.
I noticed first time I did the update there were no GPU tasks still. I forgot to turn my GPU tasks back on! (because they were always failing) which I now did.
Thanks again
Platform: Win 10 Pro 64-bit,
)
Platform: Win 10 Pro 64-bit, Xeon E3-1245 @3.90 GHz, 32 GB DDR4 ECC Ram @ 2400
Graphics: AMD Radeon RX 580, 8192 GB memory; 2 x NVidia Tesla K80
All tasks that use the
client are failing in the first few seconds. The Version 1.22 client worked well, but after downloading the 1.28 client, no WU has completed. Same graphics card is processing Milkyway@home WUs in 87 seconds while only using 19% CPU. NVidia Tesla K80 GPU Accelerator is processing same type of E@H WUs in about 36 minutes each using 99+% CPU. The NVidia documentation says the double precision arithmetic is pretty good on the K80s. Why can't we dispense with all the CPU use?
Disable beta tasks. The 1.28
)
Disable beta tasks. The 1.28 ATI/AMD app doesn’t work with cards older than Navi. (Your RX580 is incompatible with the new beta app). Turning off beta tasks will allow you to get the 1.22 app again.
_________________________________________________________________________
My GPU tasks were all
)
My GPU tasks were all failing, causing the video to crash - the screen going blank. Initially Dell replaced the NVIDIA video card, but that didnt resolve the issue - display going blank and failing the built-in 'stress test': They came out a second time and replaced the motherboard, which fixed the problem.
Julian Pedley BSc, MSc
Julianrani-[at]-gmail.com
UK, Nottinghamshire