MVA-Based Probabilistic Model of Shared Memory with a Round Robin Arbiter for Predicting Performance with Heterogeneous Workload

Ryo Kawahara, Kouichi Ono, Takeo Nakada

Abstract: Memory access contention can be a cause of performance problems and should be assessed at early stages of development. We devised a probabilistic model of shared memory for performance estimation. The calculation time is polynomial in the number of processors. The model is applicable for the region of high and heterogeneous bandwidth utilization. A round-robin arbiter is modeled using Mean Value Analysis (MVA) based approximations and incorporating non-linear dependence to the bandwidth utilization. To evaluate our model, estimated execution time is compared with the measured execution time of benchmark programs with memory access contention. We find a maximum error of 4.2% for the round-robin arbitration when we compensate for the burstiness of accesses.