Multi-GPU systems and Unified Virtual Memory for scientific applications: The case of the NAS multi-zone parallel benchmarks

• Multi-GPU and Unified Memory implementation of the Multi-Zone NAS Benchmarks. Analysis programmability performance effects Memory. per-GPU allocation have similar programming efforts. Unified-Memory version outperforms manual from 1.1x to 1.85x. GPU-based computing systems become a widely accepted solution for high-performance-computing (HPC) domain. GPUs shown highly competitive performance-per-watt ratios can exploit an astonishing level parallelism. However, exploiting peak such devices is challenge, mainly due combination two essential aspects multi-GPU execution: memory work distribution. determines data mapping GPUs, therefore conditions all distribution schemes communication phases in application. Virtual simplifies codification allocations, but its on depend how used by devices' driver going orchestrate transfers across system. In this paper we present (UM) Parallel Benchmarks which alternate computation offering opportunities overlap these phases. We analyse introduction UM support. Our experience shows that efforts introducing are those having per GPU. On evaluation environment composed 2 x IBM Power9 8335-GTH 4 GPU NVIDIA V100 (Volta), our UM-based parallelization versions 1.10x improvements sensitive information forwarded describing most convenient location specific regions. terms relationship between computational applications.