Building The 3D Pipeline ZINC22
The 3D pipeline is a collection of scripts and software packages that enable the massively parallel creation of dockable 3D molecules.
- Chemaxon License (email John for license details)
- Openeye License (email John for license details)
- Corina License (comes packaged with the executable, so you will need to provide the executable)
- SGE or SLURM queueing system installed on your cluster
- A networked file system installed on your cluster
- Clone the slurm-2 branch of the zinc-3d-build-3 repository (https://github.com/btingle/zinc-3d-build-3)
- BKS Users: Copy the software distribution from our cluster at /nfs/home/xyz/soft/*.tar.gz to the $HOME/soft directory of the user that will be running the script
- Others: Copy the software distribution from our cluster, minus corina, which you will need to supply yourself.
- When this script is submitted to a machine for the first time, it will copy and install the necessary software from $HOME/soft to local storage
- Because of this, it is important that your $HOME be global if you are running the script unmodified
- If you don't have a global home, or otherwise want your software to reside somewhere else, you can specify an alternative global directory to copy the software from by exporting SOFT_HOME prior to running submit-all-jobs.bash (see below)
- Copy your licenses into your $HOME. Copy your corina distribution (as a tar.gz) into $HOME/soft if not already present.
- This script assumes that the jchem and openeye licenses will be named ".jchem-license.cxl" and ".oe-license.txt" respectively. Corina is assumed to be a tarball named "corina.tar.gz" which contains a single directory, "corina" with executables etc.
- BKS Users: You can copy these licenses /nfs/home/xyz to your own $HOME.
Running The Script
FYI: You MUST be in the pipeline bin directory (zinc-3d-build-3) when running the submit-all-jobs script
You should also run this command in a screen, as it needs to persist until all jobs are submitted.
export INPUT_FILE=$YOUR_INPUT_FILE.smi export OUTPUT_DEST=$SOME_GLOBAL_DIRECTORY # if you don't have a global home directory (but some other global directory available), specify an alternate SOFT_HOME # export SOFT_HOME=$YOUR_SOFT_HOME export LINES_PER_BATCH=50000 export LINES_PER_JOB=50 export SHRTCACHE=/dev/shm export LONGCACHE=/tmp export BUILD_DIR=build_dir_$(whoami) ./submit-all-jobs-<slurm/sge>.bash
export INPUT_FILE=example.smi export OUTPUT_DEST=/nfs/exb/zinc22/tarballs export SBATCH_ARGS="--time=02:00:00" export LINES_PER_BATCH=50000 export LINES_PER_JOB=50 export SHRTCACHE=/dev/shm export LONGCACHE=/dev/shm export BUILD_DIR=build_dir_$(whoami) ./submit-all-jobs-slurm.bash
export INPUT_FILE=example.smi export OUTPUT_DEST=/wynton/group/bks/zinc22 export QSUB_ARGS="-l s_rt=00:28:30 -l h_rt=00:30:00 -r y" export LINES_PER_BATCH=50000 export LINES_PER_JOB=50 export LONGCACHE=/scratch export SHRTCACHE=/scratch export BUILD_DIR=build_dir_$(whoami) ./submit-all-jobs-sge.bash
export INPUT_FILE=example.smi export OUTPUT_DEST=$SCRATCH/zinc22 export SBATCH_ARGS="--cpus-per-task=1 --time=02:00:00 --requeue -q shared -C haswell" export LINES_PER_BATCH=20000 export LINES_PER_JOB=50 export MAX_BATCHES=10 export BUILD_DIR=build_dir_$(whoami) # $SCRATCH on Cori can be used for debugging, but production batch jobs should take advantage of a burst buffer allocation # this requires some custom modification of the script # export SHRTCACHE=$SCRATCH # export LONGCACHE=$SCRATCH ./submit-all-jobs-slurm.bash
The input .smi file to be built. This file should contain only two columns of data: (SMILES, NAME) with no header.
The base directory for output to be stored. The script will create a sub-directory here named $INPUT_FILE.batch-3d.d
Within this output directory there are 3 sub-directories:
In contains the input file split into fragments and sub-fragments. By default the script first splits the input file into batches of 50000, then splits those batches into sub-batches of 50. Each individual job works on one of these sub-batches. Each array batch job works on one of the batches of 50000. All of the other directories alongside 'in' share the same directory structure.
Log contains log messages from the jobs. If you are re-submitting a file, be aware that log messages from previous runs on this file will be overwritten.
Out contains tar.gz output from each job. The tarballs should contain a number of 3d molecule formats for each molecule in the batch, including 1 or more db2.gz files.
The base working directory for the script. By default it is /tmp. If you want your jobs to run a little faster and are not concerned about memory usage, you can set the working directory to /dev/shm.
How many lines of the source .smi file should be processed per array batch job
How many lines of the batch .smi file should be processed per array task
Additional arguments for the sbatch command. It is recommended to set a --time limit, as build jobs will save progress & terminate if they are still running two minutes before the --time limit.
--requeue allows jobs that reach the time limit to run later, utilizing their saved progress.
Slurm seems to not send signals to our jobs correctly, so this feature should be ignored on slurm clusters. (for now)
Additional arguments for the qsub command. Similar to slurm, it is recommended to set a time limit, but you will need to manually specify both s_rt & h_rt. In the example, we set s_rt to be a minute and thirty seconds before h_rt. s_rt is the point where jobs will save progress and terminate, h_rt is when they will be forcibly terminated, even if they've not finished saving.
-r y allows jobs that reach the time limit to run later, utilizing their saved progress.
Each batch job will contain LINES_PER_BATCH/LINES_PER_JOB jobs, and there will be a maximum of MAX_BATCHES batches submitted at any given time. By default this value is 30, which corresponds to 30,000 queued jobs at any given time if there are 1000 jobs per batch.
At BKS, we currently store the tarred output of the pipeline @ /nfs/exb/zinc22/tarballs. Currently, we use the following command to repatriate output from other clusters to our cluster:
### migrate_output.bash for output in $OUTPUT_DEST/*.batch-3d.d; do echo "starting rsync on $output to $MIGRATE_USER@files2.docking.org" sshpass -f $PW_FILE rsync -arv $output/out $MIGRATE_USER@files2.docking.org:/nfs/exb/zinc22/tarballs/$(basename $output).out done
sshpass is optional here but preferable for convenience's sake. Since files2.docking.org is only visible within the UCSF network, any clusters outside will need to maintain a network tunnel when rsyncing.
Sometimes an output tarball will have few or no entries within. Certain molecule types [elaborate] will fail to be built, and often these molecules get bunched together (i.e if the input file is sorted by SMILES). Additionally, a small percentage of all molecules may fail to be processed by corina or amsol. If neither of these explain what is causing your missing entries, check that tarball's corresponding log entry for more info.
This script was designed for a 64 bit architecture. You will likely run into some library issues trying to run it on 32 bit machines. If that's all you have you can try to swap out libg2c.so.0* in lib.tar.gz with a 32 bit version, but I cannot help you beyond that.
It is safe to re-run the same file multiple times- the script takes care of making sure not to re-run any jobs that have already completed successfully prior. This is only the case if that file's corresponding batch-3d.d output directory has not been moved or deleted.
For example, if one of your nodes went down and caused a bunch of jobs to fail, it would be safe to re-run ./submit-all-jobs.bash to re-submit those jobs. (assuming there are no jobs for that file currently queued/running)