An Analytical Model to Predict Performance Impact of DRAM Bank Partitioning
Dong Uk Kim, Seokju Yoon, Jae Wook Lee · 2013
The main memory system is one of the most important shared resources in multicore platforms. To maximize bandwidth and minimize latency, modern DRAM systems exploit spatial locality in a memory access stream. However, multiple memory request streams from multiple concurrent threads are interleaved, hence killing the spatial locality of DRAM accesses. This increases bank conicts and decreases row buer hit rates. To address this problem, we advocate workload-aware, non-uniform DRAM bank partitioning to improve overall system throughput and fairness. To make best use of a limited number of DRAM banks oered by multicore platforms, we propose an analytical approximation model that predicts the program slowdown when the number of allocated banks is varying and apply it to nd an optimal bank allocation for a given multi-programmed workload. Our preliminary evaluation shows promising results| the non-uniform bank partitioning guided by the prediction model improves overall system throughput by up to 3% (and 0.8% on average) and fairness by up to 26% (and 7.5% on average) over static uniform partitioning.