We have started getting this error message when running various jobs (see
the example screenshot below from a blob picking job).
After the cryosparc job fails, I get this message from systemlogd:
kernel:[253932.878477] watchdog: BUG: soft lockup - CPU#2 stuck for 48s! [gnome-shell:1808] (this message keeps repeating with a different CPU# until the system is physically rebooted).
We are not sure what the problem is… Some googling led me to believe it was something to do with the swapfile. However, deleting or replacing it had no effect. Also, if I monitor the system with htop while the job is running I notice the memory cache slowly increases until it reaches the maximum at which point the job crashes…
If anyone has encountered this before or any ideas on what to do, it would be very helpful!
@BWise Which other job types gave you this same
Is there anything remarkable about the input data, such as the size of the structure determination target, the data format or the data quality?
Please can you also provide:
- GPU model(s) and driver version (
- Linux kernel (
- DRAM size (
- worker CUDA version:
eval $(cryosparc_worker/bin/cryosparcw env)
python -c "import pycuda.driver; print(pycuda.driver.get_version())
Hi, thanks for the reply!
No, I don’t think there is anything remarkable about this data. The target is about 120 kDa, and 200-300 angstoms in size, the raw movies are in .tiff format. It is a new target and I just recently collected a first dataset of about 8000 movies. I followed my usual workflow: import movies → patch motion → patch CTF → curate exposures, all without problems. The error appeared after.
Error occurs so far during Blob picking and Extract from micrographs.
cryosparc is up to date with the latest patches.
GPU model(s): 2xQuadro RTX5000 and driver version: 495.29.05
Linux kernel: 5.13.0-30-generic
worker CUDA version: (11, 5, 0)
total used free shared buff/cache available
Mem: 251 2 150 0 98 247
Swap: 255 0 255
thanks for your help!
@BWise As a first step in troubleshooting this issue, please configure your worker and test the failed job with CUDA-11.2:
I followed your advice and configured the worker to use CUDA-11.2. I cloned and reran the extraction job. It worked!!
Again I monitored the system with htop while the job was running and although the memory cache reached the maximum like last time, the job did not crash. 2D classification is running now, so far without any problems. So hopefully that solves the problem
One thing though: when I tried to change the CUDA path with cryosparcw i get an error “cryosparcw: command not found”. Am I missing something?
For this I used nano to edit the cryosparc_worker/config.sh file, but for future reference it would be useful to know why the cryosparcw command does not work for me.
<path-to-cryosparc-worker-software>/bin has been added to the
PATH environment variable, you should run (based on the output you posted earlier):
/usr/local/cryosparc/cryosparc/cryosparc_worker/bin/cryosparcw newcuda <cuda-path>
that works, thanks!
Hi again, Unfortunately I spoke too soon.
2D classification ran without problems.
However, the exact same error described above occurred again when extracting the same particles with a smaller box size.
originally with CUDA 11.5 the extraction job would consistently crash around 800 micrographs out of 4000. Now with CUDA 11.2 the job crashed around 3300 out of 4000. Why it completed successfully the first time with a larger boxsize. I do not know…
Other ideas? thanks again!
@BWise Unfortunately, I don’t have any suggestions from the software side at this time. Are you confident in the integrity of your GPU hardware, i.e., non of the GPUs is failing?
While monitoring the GPUs with nvtop during a running job I did indeed notice one of the GPUs was exhibiting strange behavior. So, a failing GPU seems to be the culprit
So far I have not seen this error anymore if I avoid using this failing GPU