Unix I/O performance in workstations and mainframes

Abstract: Rapid advances in processor performance have shifted the performance bottleneck to I/O sys-tems. The relatively slow rate of improvement in I/O is due in part to a lack of quantitative performance analysis of software and hardware alternatives. Using a new self-scaling I/O benchmark, we provide such an evaluation for 11 hardware configurations using 9 variations of the Unix operating system. In contrast to processor performance comparisons, where factors of 2 are considered large, we find differences of factors of 10 to 100 in I/O systems. The principal performance culprits are the policies of different Unix operating systems; some policies on writes to the file cache will cause processors to run at magnetic disk speeds instead of at main memory speeds. These results suggest a greater emphasis be placed on I/O performance when making operating system policy decisions