Files
merlin-cryosparc/README.md

2.5 KiB

CryoSparc Merlin scripts

This repository collects cryosparc 'lanes' suitable for running jobs at PSI.

Lanes

  • gpu. Default lane. 7 day time limit.
  • gpu-big. Requests requests GPUs with 11GB of video memory, plus 100G memory
  • gpu-short. Short requests. 2 hour time limit.
  • gpu-rtx2080ti. Specialty queue for jobs that benefit from the RTX 2080Ti cards
  • cpu-daily. CPU-only lane. 1 day time limit.

Installing

This repo can be cloned using git on merlin. Typically this is done from your cryosparc install directory (/data/user/<user>/cryosparc/), although this isn't required.`

git clone https://git.psi.ch/lsm-hpce/merlin-cryosparc.git config

Next, all cluster_info.json scripts need to be modified to include the correct worker path by replacing USERNAME. A script is provided to do this automatically:

dev/install_filters.sh

Check that the cuda version in cluster_script.sh is correct. If not, update the files as described in 'Change CUDA versions' below.

Finally, connect the newly modified scripts to cryosparc. This should be done on same machine cryosparc runs on. To connect all lanes:

./connect_all.sh

Alternately, you can install individual lanes manually:

cd <lane>
cryosparcm cluster connect

Updating

Scripts are updated regularly with features and bug fixes. To get the newest version, update git:

git pull

(Optional) If you want to remove old lanes (e.g. when updating from scripts v1.5.0 to v2.0.0), run

./remove_old_lanes.sh

Then, connect the lanes to your cryosparc cluster as in installation:

./connect_all.sh

Change CUDA versions

The scripts load cuda/10.0.130 by default. Newer cryosparc versions may require using a newer cuda version. This can be changed by either loading a different cuda module while installing/upgrading the worker, or using the following command:

module load cuda/10.0.130
$CRYOSPARC_HOME/cryosparc2_worker/bin/cryosparcw newcuda $CUDA_PREFIX

The submission scripts then need to be updated to match this cuda version:

module load cuda/10.0.130
sed -ri "s|cuda/\S+|cuda/$CUDA_VERSION|" */cluster_script.sh
./connect_all.sh

Developers

If you plan on committing changes to this repository, make sure you run dev/install_filters.sh to install git hooks and filters.

When releasing a new version, dev/mkrelease.sh will update version numbers consistently.

References