An analysis of the I/O performance of the OpenPower 710 on Linux

The importance of a well tuned I/O subsystem is beyond doubt. In many cases, I/O may be the real bottleneck of an application and understanding its limitations and capabilities may be of interest both from the research and development points of view. In this paper we analyze the I/O subsystem of an IBM e-server OpenPower 710 with 4 SCSI disks for asynchronous and OS unbuffered I/O. The analysis is done with the aid of IOAgent, a tool that allows to generate synthetic I/O workloads for Linux. With the performance measures we build a statis-tical model of the I/O subsystem behavior. The most relevant results show that the number of disks being accessed has a positive influence on the abso-lute performance of an I/O bound simulated appli-cation. However, in general, our results also show that sequential I/O activity benefits more from a small number of processes and application buffers, while random I/O activity benefits more from a large number of processes and application buffers