search

docking-org / pydock3

Python package wrapping the DOCK Fortran program and providing several tools built on top of it.
Other
11 stars 3 forks source link

DockOpt-issued docking jobs occasionally run out of memory #28

Open ianscottknight opened 1 year ago

ianscottknight commented 1 year ago

Credit to @gregorpatof for bringing this to my attention.

DockOpt-issued docking jobs occasionally run out of memory (e.g., witnessed on n-1-16 on Gimel while running other CPU-intensive jobs on same machine). Memory requirement should be specified to job scheduler being used (e.g., --mem flag for sbatch Slurm command). @jir322 has reported that it was empirically found that 1 GB was found to be sufficient when 500 MB proved insufficient (on AWS, I believe).

In general, more granular control over the behavior of the job scheduler as determined by the parameters of the batch submit command should be made available to the user. This itself should probably become an enhancement issue.

jir322 commented 1 year ago

All is takes is one "cheating" job to not properly reserve memory, to allow a machine to overload memory and fail as you describe.

I wonder if the default memory request for each job is 0 and not 1 GB. Suggest changing default memory requirement to 1 GB (or even 1.5 or 2.0 GB) . So if you forget to assign a memory footprint, a reasonable default is supplied.

Perhaps this will help?

John Irwin UCSF Pharmaceutical Chemistry http://irwinlab.compbio.ucsf.edu

On Wed, Mar 29, 2023 at 2:15 PM Ian Scott Knight @.***> wrote:

Credit to @gregorpatof https://github.com/gregorpatof for bring this to my attention.

DockOpt-issued docking jobs occasionally run out of memory (e.g., witnessed on n-1-16 on Gimel while running other CPU-intensive jobs on same machine). Memory requirement should be specified to job scheduler being used. (e.g., --mem flag for sbatch Slurm command). @jir322 https://github.com/jir322 has reported that it was empirically found that 1 GB was found to be sufficient when 500 MB proved insufficient (on AWS, I believe).

In general, more granular control over the behavior of the job scheduler as determined by the parameters of the batch submit command should be made available to the user. This itself should probably become an enhancement issue.

— Reply to this email directly, view it on GitHub https://github.com/docking-org/pydock3/issues/28, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABIR2H4E3CUXW54NUY3TJF3W6SQ5ZANCNFSM6AAAAAAWMNR4LE . You are receiving this because you were mentioned.Message ID: @.***>