Tuning the performance of I/O-intensive parallel applications

Abstract

Getting good I/0 performance from parallel programs is a critical problem for many application domains. In this paper, we report our experience tuning the I/0 performance of four application programs from the areas of satellite-data processing and linear algebra. After tuning, three of the four applications achieve application-level 1/0 rates of over 100 MB/s on 16 processors. The total volume of 1/0 required by the programs ranged from about 75 MB to over 200 GB. We report the lessons learned in achieving high I/0 performance from these applications, including the need for code restructuring, local disks on every node and knowledge of future I/0 requests. We also report our experience on achieving high performance on peer-to-peer configurations. Finally, we comment on the necessity of complex I/0 interfaces like collective I/0 and strided requests to achieve high performance.