Intelligent colocation of HPC workloads

Many HPC applications suffer from a bottleneck in the shared caches, instruction execution units, I/O or memory bandwidth, even though remaining resources may be underutilized. It is hard for developers and runtime systems to ensure that all critical are fully exploited by single application, so an attractive technique increasing system utilization colocate multiple on same server. When share resources, however, contention lead reduced application performance. In this paper, we show server efficiency can improved first modeling expected performance degradation of colocated based measured hardware counters, then exploiting model determine optimized mix applications. This paper presents new intelligent resource manager makes following contributions: (1) machine learning predict counters (2) scheduling scheme deployed existing enable co-scheduling with minimum degradation. Our results our approach achieves improvements 7 % (avg) 12 (max) compared standard policy commonly used job managers. • colocating jobs machine. Jobs contending experience Hardware characterize resource. Greedy strategy improve batch systems.