CSC Computing Environment

Using HyperQueue for farming Gaussian jobs on Puhti

This tutorial requires that

This tutorial is done on Puhti


💬 HyperQueue is a tool for efficient sub-node task scheduling and well suited for farming and running embarrassingly parallel jobs.

💬 In this example, we have several similar molecular structures and would like to know how they differ energetically.

The workflow of this exercise

  1. Download 200 sample molecular structures
  2. Convert these structures to Gaussian format
  3. Construct the corresponding Gaussian input files
  4. Build a sbatch-hq command list to run the jobs
  5. Submit the job using the sbatch-hq wrapper
  6. Analyze the results

Download 200 sample 3D molecular structures

  1. Create and enter a suitable scratch directory on Puhti (replace <project> with your CSC project, e.g. project_2001234):
mkdir -p /scratch/<project>/$USER/gaussian-hq
cd /scratch/<project>/$USER/gaussian-hq
  1. Download the 200 C7O2H10 structures that have originally been obtained from the QM9 dataset:
  1. Unpack the archive:
tar -xzf C7O2H10.tar.gz
  1. Go to the directory containing the structure files that are in .mol format:
cd C7O2H10

Convert the structures to Gaussian format

💬 Gaussian is a program for molecular electronic structure calculations.

  1. Use OpenBabel to convert the structures to Gaussian format:
module load openbabel
obabel *.mol -ocom -m
  1. Now we have converted the 200 structures into .com format that is used by Gaussian.

Construct the corresponding Gaussian input files

💬 In this example we want to do a b3lyp/cc-pVDZ calculation on these structures, i.e. a hybrid density functional theory calculation using the B3LYP exchange-correlation functional and the cc-PVDZ basis set.

  1. Add the b3lyp/cc-pVDZ keyword at the beginning of each .com file:
sed -i '1s/^/#b3lyp\/cc-pVDZ \n/' *.com
  1. Set 4 cores per job by adding the flag %NProcShared=4 to each input file:
sed -i '1s/^/%NProcShared=4\n/' *.com
  1. Now you have 200 complete Gaussian input files corresponding to the original molecular structures and the method of choice.

Build a command list to run the jobs as a HyperQueue task array

💬 A task array can sometimes be lengthy so rather than typing it by hand it is more feasible to use bash scripting to create a suitable task list file for HyperQueue.

  1. Move back up to your main directory:
cd ..
  1. Create the task list and name it commandlist:
for f in ${PWD}/C7O2H10/*.com; do echo "g16 < $f >> output/$(basename ${f%.*}).log" >> commandlist; done
  1. Inspect the task list with more, less or cat. The file should look like:
g16 < /scratch/<project>/$USER/gaussian-hq/C7O2H10/ >> output/dsC7O2H10nsd_0001.log
g16 < /scratch/<project>/$USER/gaussian-hq/C7O2H10/ >> output/dsC7O2H10nsd_0002.log
g16 < /scratch/<project>/$USER/gaussian-hq/C7O2H10/ >> output/dsC7O2H10nsd_0003.log
  1. Notice that the output will be directed into a directory called output. Create this directory:
mkdir -p output

Run the HyperQueue task array with sbatch-hq

💬 Running a HyperQueue task array is similar to running a Slurm array job. However, HyperQueue packs the individual tasks within a single Slurm job step and is thus much more efficient, especially if there are a huge number of tasks. In this case, submitting the job is also very easy since we can use the sbatch-hq wrapper to avoid having to create a batch script by hand.

  1. Submit the list of Gaussian commands using sbatch-hq:
module load sbatch-hq gaussian
sbatch-hq --cores=4 --nodes=1 --account=<project> --partition=small --time=00:15:00 commandlist

💬 The sbatch-hq command creates and submits a batch script that starts the HyperQueue server and worker(s) and submits the task array with inputs read from the commandlist file. The following resources are requested:

💬 Given that 40 cores are requested for running 200 tasks, each using 4 cores, 10 tasks are able to run concurrently. The number of commands in the file can (usually should) be much larger than the number of commands that can fit running simultaneously on the reserved resources to avoid creating too short Slurm jobs.

Monitor the job

  1. You can monitor the Slurm queue with (replace <slurmjobid> with the assigned Slurm job ID):
squeue -j <slurmjobid>
# or
squeue --me
# or
squeue -u $USER
  1. This does, however, not provide you information about the progress of the individual sub-tasks. To monitor these, export the location of the HyperQueue server and use the hq commands:
export HQ_SERVER_DIR=$PWD/hq-server-<slurmjobid>   # replace <slurmjobid> with the actual id of your Slurm job
hq job info 1
  1. Once the workflow has finished (should take a bit more than 10 minutes), print a list of the b3lyp/cc-pVDZ energies for each of the 200 structures sorted by energy (most stable structure first):
grep -r "E(RB3LYP)" output | sort -k6 -n -o energies.txt
  1. Using head energies.txt, the output should look like:
output/dsC7O2H10nsd_0015.log: SCF Done:  E(RB3LYP) =  -423.218630672     A.U. after   14 cycles
output/dsC7O2H10nsd_0192.log: SCF Done:  E(RB3LYP) =  -423.216601925     A.U. after   12 cycles
output/dsC7O2H10nsd_0193.log: SCF Done:  E(RB3LYP) =  -423.214963908     A.U. after   12 cycles
output/dsC7O2H10nsd_0028.log: SCF Done:  E(RB3LYP) =  -423.214781165     A.U. after   13 cycles
output/dsC7O2H10nsd_0037.log: SCF Done:  E(RB3LYP) =  -423.214421420     A.U. after   14 cycles
output/dsC7O2H10nsd_0026.log: SCF Done:  E(RB3LYP) =  -423.214326717     A.U. after   14 cycles
output/dsC7O2H10nsd_0008.log: SCF Done:  E(RB3LYP) =  -423.213824577     A.U. after   14 cycles
output/dsC7O2H10nsd_0036.log: SCF Done:  E(RB3LYP) =  -423.212123483     A.U. after   14 cycles
output/dsC7O2H10nsd_0025.log: SCF Done:  E(RB3LYP) =  -423.212093937     A.U. after   14 cycles
output/dsC7O2H10nsd_0191.log: SCF Done:  E(RB3LYP) =  -423.211777369     A.U. after   13 cycles