Improving UNIX kernel performance using profile based optimization

Several studies have shown that operating system performance has lagged behind improvements in application performance. In this paper we show how operating systems can be improved to make better use of RISC architectures, particularly in some of the networking code, using a compiling technique known as Profile Based Optimization (PBO). PBO uses profiles from the execution of a program to determine how to best organize the binary code to reduce the number of dynamically taken branches and reduce instruction cache misses. In the case of an operating system, PBO can use profiles produced by instrumented kernels to optimize a kernel image to reflect patterns of use on a particular system. Tests applying PBO to an HP-UX kernel running on an HP9000/720 show that certain parts of the system code (most notably the networking code) achieve substantial performance improvements of up to 35% on micro benchmarks. Overall system performance typically improves by about 5%.