Object-centric bank partition for reducing memory interference in CMP systems

Abstract

This work introduces a novel object-centric bank partition (OBP) to mitigate both the inter-thread and intra-thread interference. The key idea is to break bank sharing relationship among the simultaneously accessed data objects, instead of only focusing on the co-running threads. At sampling runs, we profile each thread to identify the simultaneously accessed objects. At actual runs, using the profiling information, the operating system partition banks at both the thread and object level. We have implemented OBP in Linux 2.6.32 kernel and evaluated its benefits on real machines. Experimental results show that OBP achieves an encouraging enhancement it terms of performance.