Analysis and Modeling of Resource Contention Effects based on Benchmark Applications

Abstract

Benchmark suites, such as SPLASH-3, can be used to investigate the parallel runtime of complex applications from science and engineering on modern multicore platforms. In high performance computing (HPC), a high number of such parallel tasks might be executed concurrently. Depending on the specific operations performed by the tasks, their runtimes are often affected by contention for hardware resources, such as communication networks, the main memory, or hard disks. In this article, we investigate the effects of resource contention for the concurrent execution of application and kernel tasks of the SPLASH-3 benchmark suite. The parallel tasks are executed on a heterogeneous HPC cluster and an approach for modeling the measured runtimes without and with resource contention is presented. The proposed modeling is used for runtime predictions and the achieved accuracy is evaluated. To reduce the time required for executing multiple parallel tasks of the SPLASH- 3 benchmark suite, the runtime prediction is integrated into a scheduling method for parallel tasks on heterogeneous HPC clusters. Performance results are shown to demonstrate the improvements achieved with a better utilization of parallel and concurrent executions.