Workload Aware Utilization Optimization for a Petaflop Supercomputer: Evidence Based Assessment Using Statistical Methods

Abstract

Nowadays, computing resources like supercomputers are shared by many users. Most systems are equipped with batch systems as their resource managers. From a user's perspective, the overall turnaround of each submitted job is measured by time-to-solution which consists of the sum of batch queuing time and execution time. On a busy machine, most jobs spend more time waiting in the batch queue than their real job executions. And rarely this is a topic of performance tuning and optimization of parallel computing. we propose a workload aware method systematically to predict jobs' batch queue waiting time patterns. Consequently, it will help user to optimize utilization and improve productivity. With workload data gathered from a supercomputer, we apply Bayesian framework to predict the temporal trend of long-time batch queue waiting probability. Thus, the workload of the machine not only can be predicted, we are able to provide users with a monthly updated reference chart to suggest job submission assembled with better chosen number of CPU and running time requests, which will avoid long-time waiting in batch queue. Our experiment shows that the model could make over 89% correct predictions for all cases we have tested.