An Online Cache Miss Rate Curves Generating And Shared Cache Partition Method

Shared caches in multi-core processors play important roles such as acceleratinginter-core communication. Yet the inherent cache contention by multiple processes onsuch architectures causes a significant performance degrading. To address this problem,cache partition has been studied based on the prediction of the cache miss rate curve(MRC) of the concurrently running programs. Online MRC estimation, however, eitherrequires special hardware support or incurs a high overhead when conducted purely insoftware.This paper presents a pure software, low overhead, and high accurate MRCgeneration method. Cache is partitioned based on MRC. Cache accesses behavior ofprograms is fractal. Online MRC prediction method based on fractal is proposed andhence called FractalMRC. This method exploits the hardware performance monitoringcounters available on most modern processors to record runtime information, so changesto the applications, or hardware are not required. This method can be used on real systemswithout additional hardware. Parameters of prediction model can be calculated usingruntime information. Then parameters are feed to fractal prediction model to obtain cachemiss rate curve. MRC is provided to efficiently partition shared cache in OS level,improving the overall system performance.We compare our online prediction with the real measurement for26applicationstaken from SPEC CPU2006benchmark suit. On average, the prediction error achieves3.13%. This method requires tracing period of420ms, and subsequent24us to process thedata. Finally, we apply our prediction results to partition the shared cache system, andresults show that FractalMRC can be applied to size cache partitions on real systems toimprove performance by up to20%.