Characterizing Shared Memory Multiprocessor Benchmarks for Future Chip Multiprocessor Architectures Using Instruction Flow Analysis

Abstract

For forty years, transistor counts on integrated circuits have doubled roughly every two years, enabling computer architects to double the clock speed of processors. Recently, heat dissipation and power consumption trends have forced chip designers to add larger caches and more cores per chip, instead of increasing clock speed with the extra transistors. This has provided challenges for programmers who wish to continue increasing application performance as though the speed of a uniprocessor had continued doubling. In this characteristic study, we examine the effect of the operating system on a set of parallel benchmarks run on a simulated many-core processor. Past research has shown that the performance of the OS code has a large impact on application performance; however, most studies ignore the OS and focus on the application code. This work will characterize performance bottlenecks and show possible areas that could be improved. We found that resource contention in the kernel was limiting the efficiency of the benchmarks.