Latencies of Conflicting Writes on Contemporary Multicore Architectures

Abstract. This paper provides a detailed investigation of latency penalties caused by repeated memory writes to nearby memory cells from different threads in parallel programs. When such writes map to the same corresponding cache lines in multiple processors, one can observe the so called false sharing effect. This effect can unnecessarily hamper parallel code due to the line granularity based cache hierarchy, which is common on contemporary processor architectures. In this contribution, a benchmark allowing for quantitative estimates about the consequences of the false sharing effect, is presented. Results show that multicore architectures with shared cache can reduce unwanted effects of false sharing