Queue Managers for High-Performance Computing

High-performance computing (HPC) clusters are designed to complete highly-parallel tasks in a short time. Properly leveraging such clusters requires utilizing large numbers of compute nodes at the same time, which requires special configurations for the QCFractal manager. This part of the guide details several routes for configuring HPC clusters to use either large numbers tasks that each use only a single node, or deploying a smaller number of tasks that use multiple nodes.

Note: This guide is currently limited to using the Parsl adapter and contains some configuration options which do not work with other adapters.

Many Nodes per Job, Single Node per Application

The recommended configuration for a QCFractal manager to use multi-node Jobs with tasks limited to a single node is launch many workers for a single Job.

The Parsl adapter deploys a single manager'' per Job and uses the HPC systems's MPI task launcher to deploy the Parsl executors on to the compute nodes. Each "executor" will run a single Python process per QCEngine worker and can run more than one worker per node. The ``manager will run on the login or batch node (depending on the cluster’s configuration) once the Job is started and will communicate to the workers using Parsl’s ØMQ messaging protocol. The QCFractal QueueManager will only need to Parsl manager and not any of the workers.

See the example page for details on how to configure Parsl for your system. The configuration setting common.nodes_per_job defines the ability to make multi-node allocation requests to a scheduler via an Adapter.

Many Nodes per Job, More than One Node per Application

The recommended configuration for using node-parallel tasks is to have a single QCFractal worker running on the batch node, and using that worker to launch MPI tasks on the compute nodes. The differentiating aspect of deploying multi-node tasks is that the QCFractal Worker and QCEngine Python process will run on different nodes than the quantum chemistry code.

The Parsl implementation for multi-node jobs will place a Parsl single executor and interchange on the login/batch node. The Parsl executor will launch a number of workers (as separate Python processes) equal to the number of nodes per Job divided by the number of nodes per Task. The worker will call the MPI launch system to place quantum-chemistry calculations on the compute nodes of the clusters.

See the example page for details on how to configure Parsl for your system.