23rd Euromicro International Conference on Parallel, Distributed, and Network-Based Processing (PDP 2015), 4–6 March 2015, Turku, Finland
Abstract—This work studies utilization of shared caches byapplications running concurrently on different cores ofmulticore systems. Knowledge about program contention dueto shared resources is important for various design problemsconcerning multicore architectures. It is needed for powerestimation, scheduling of parallel applications and design ofshared memories. Moreover, deep understanding of programsbehavior is especially needed for the development of accuratemodels that are able to predict misses caused by sharedresources in multicore systems.We present a methodology that is able to examine theinteraction of applications in shared caches. Our experimentsshow a positive impact of data sharing by minimizing misses inshared L2 caches over a wide range of L2 cache sizes forapplications from the Mediabench suite. Up to 25% lowermisses in the last level cache can be observed for embeddedapplications, when data are allowed to be shared amongprograms running on different cores.