Improving DRAM Bandwidth Utilization with MLP-Aware OS Paging

Abstract

Optimal use of available memory bank-level parallelism and channel bandwidth heavily impacts the performance of an application. Research studies have focused on improving bandwidth utilization by employing scheduling policies and request re-ordering techniques at the memory controller. However, potential to extract memory performance by intelligent page allocation that maximizes opportunity for bank-level parallelism and row buffer hits is often overlooked. The actual physical page location in memory has a huge impact on bank conflicts and potential for prioritizing low-latency requests such as row buffer hits. We demonstrate that with more intelligent virtual to physical paging mechanisms it is possible to reduce bank conflicts at the memory and achieve higher bandwidth utilization. Such intelligent paging mechanisms can then form a basis for other request re-ordering techniques to further improve memory performance. In this study we only focus on virtual-to-physical paging techniques and demonstrate 38.4% improvement on DRAM bandwidth utilization with a profile-based scheme. We study a wide variety of workloads from varied benchmark suites. We present results for profile based as well as preliminary results for dynamically adaptive paging techniques. Our results demonstrate improved bandwidth utilization with DRAM aware page layouts. Dynamic paging schemes further demonstrate the potential of run-time adaptive techniques in improving bandwidth utilization of increasingly parallel multi-channel main memory systems.