Debugging Of Multi-core Architecture Performance

Multicore technology has become the mainstream trend of modern processors development,which brings big challenges to software design.How to develop fast and effective parallel programs to take advantage of resources on multicore platforms, is always a hard problem that both industry and research fields keep trying to resolve. Multicore performance debugging is a great facility in helping deliver high performance parallel programs.By analyzing the program behaviors,it can find out performance bottlenecks,and further effectively support performance optimization.Considering the dynamics and uncertainty of parallel programs,traditional code analyzing techniques can hardly find out performance problems.Some recent research work proposed software solutions to analyze program behaviors,which generally results in heavily slowdown and inaccurate data.Although hardware implementations could largely reduce runtime overhead,the structure extensions could hurt program's performance.In this paper,we present a performance monitor(PerfMon).By extending cache structure and cache coherence protocol,it monitors sharing,occupying and conflicting on shared cache among the computing cores.The monitored result can be feed up to software through performance counters so as to support performance analyzer and dynamic scheduler on multicore platforms.Experimental results show that PerfMon only introduces very low overhead in both hardware extension and runtime,while effectively helps scheduler archive good speedup in average case.