Jole: a library for dynamic job-level parallel workloads

Abstract

Problems in scientific computing often consist of a workload of jobs with dependencies between them. Batch schedulers are job-oriented, and are not well-suited to executing these workloads with complex dependencies. We introduce Jole, a Python library created to run these workloads. Jole has three contributions that allow flexibility not possible with a batch scheduler. First, dynamic job execution allows control and monitoring of jobs as they are running. Second, dynamic workload specification allows the creation of workloads that can adjust their execution while running. Lastly, dynamic infrastructure aggregation allows workloads to take advantage of additional resources as they become available. We evaluate Jole using GAFolder, a protein structure prediction tool. We show that our contributions can be used to create GAFolder workloads that use less cluster resources, iterate on global protein structures, and take advantage of additional cluster resources to search more thoroughly.