Development of an UltraScan Meta-Scheduler for HPC Job Submission

点击翻译按钮获取中文摘要

Abstract UltraScan is a mature, open-source multi-platform analysis software for analytical ultracentrifugation (AUC) experiments. It includes a desktop GUI system, a relational database and a web-based LIMS server, a feature rich desktop environment, including a GMP data acquisition module, and a unique-to-the field high-performance parallel computing (HPC) component. This HPC component is the focus of this supplemental request. UltraScan has been implemented on high-performance computing (HPC) clusters to take advantage of parallel computing infrastructure. UltraScan has been fully integrated into the open source Apache Airavata framework, which manages the HPC executions and data transfers for UltraScan on nine different HPC resources across the world. It is operated by the Cyberinfrastructure Integration Research Center at Indiana University to provide a grid middleware layer for UltraScan and other collaborators. Apache Airavata enables access to shared installations of the HPC analysis software on supercomputers allocated by NSF/XSEDE and institutional clusters at several universities. New systems are continually integrated as they become available. The increasing number and diversity of available supercomputing platforms that have been integrated into the UltraScan system through Apache Airavata raise several new unmet challenges in usability and scalability which we propose to address with this supplement request. In Aim 1 we propose to extend the Apache Airavata architecture by developing a new meta-scheduler to abstract the job submission to any available HPC or cloud resource by automatically determining the best available cluster and queue, so users no longer have to know how to determine this by themselves. The scheduler will determine optimal availability for jobs submitted by the UltraScan community account and will automatically optimize the submission based on the job’s requirement for memory, compute time, and compute cores. The meta-scheduler will support dynamic resubmission of jobs in case a cluster goes into maintenance, or if queuing times become prohibitive. It will also be configurable, simplifying addition of new resources. In Aim 2 we will develop a resource requirement prediction microservice. The microservice will be initially trained on metadata from previous jobs stored in all UltraScan LIMS databases. The microservice will be dynamically updated as jobs complete. An API will be exposed and utilized by the meta-scheduler to receive job- and resource-specific maximum wall times. This will reduce queuing times and improve the utilization of oversubscribed resources.