Link Search Menu Expand Document
Quill

Input files

Example input files can be found in the conf/example folder. All possible parameters in the input file are described in the quill.conf.example file.

Run

To run Quill, use the run.sh script from the root folder and pass the name of the input file as a parameter. For example,

./run.sh /home/user/my-problem

By default, run.sh relies on mpirun. Rewrite the script itself if different behavior is required.

The number of MPI threads is determined by the n_sr parameter in the input file and should never be explicitly set to a different value.

Note: In order to run Quill, a Python interpreter is required on all nodes; it is used to parse the input file.

Parallelization in Quill is optimized for processess with adjacent ranks being on the same node/socket, so mapping by socket is advised. Binding processes to cores is also advised to make sure the memory is always allocated. For OpenMPI, the corresponding configuration is mpirun --map-by socket --bind-to core. The configuration can be changed in the run.sh script.

Running with slurm

Slurm is a popular workload manager for clusters. See Slurm documentation for comprehensive details.

The most common way to schedule a Quill job is through sbatch. An example shell-script run_slurm_example.sh can be found in the root of the project. As run.sh, it accepts the path to an inpit file

sbatch -N 1 -n 32 run_slurm_example.sh /home/user/my-problem

The most commonnly used arguments taken by sbatch are:

  • -N, --nodes=: number of nodes (-N 2 for 2 nodes).
  • -n, --ntasks=: number of tasks (-n 32 for 32 tasks)
  • -o, --output=: file to redirect the process output to, e.g. slurm-%x.%j.out (refer to Slurm documentation for patterns)

Arguments can be explicitly passed to sbatch or written in the shell-script itself.

One can also use salloc and run Quill using run.sh from the interactive Slurm session.

Note: the number of allocated slurm tasks (ntasks) does not affect the number of processees spawned by Quill (determined by the n_sr parameter, see above). You have to make sure they are the same value.

Running multiple jobs based on a template

Refer to the description in the run_slurm_example.py script for running several Slurm jobs based on a input template.