Dynamic Page Policy Using Perceptron Learning.

Modern multicore processors running data-intensive applications concurrently generate memory requests that exhibit different spatial and temporal locality. These interleaved memory requests result in an increased random memory access pattern as seen by the memory controller. Different page policies have been implemented to decide if a row should be kept open or close after an access. Appropriate choice of page policy is critical in reducing the memory access latency, especially in massively parallel main memory systems intended for high performance computing. In this paper, we introduce a dynamic page policy that uses perceptron learning to make the page open/closure decision by examining a wide variety of features extracted from application's memory access behavior. Simulation over multiprogramming SPEC CPU workloads indicates that our policy significantly reduces the memory access latency by reducing the row-buffer conflicts, and improves the overall performance by 5.5% on average as compared to a previously proposed dynamic page policy that makes adaptive page open/closure decision based on the bank-level page hit rate.