Modern DRAM Memory Systems: Performance Analysis and Scheduling Algorithm

Abstract

The performance characteristics of modern DRAM memory systems are impacted by two primary attributes: device datarate and row cycle time. Modern DRAM device datarates and row cycle times are scaling at different rates with each successive generation of DRAM devices. As a result, the performance characteristics of modern DRAM memory systems are becoming more difficult to evaluate at the same time that they are increasingly limiting the performance of modern computer systems. In this work, a performance evaluation framework that enables abstract performance analysis of DRAM memory systems is presented. The performance evaluation framework enables the performance characterization of memory systems while fully accounting for the effects of datarates, row cycle times, protocol overheads, device power constraints, and memory system organizations.

This dissertation utilizes the described evaluation framework to examine the performance impact of the number of banks per DRAM device, the effects of relatively static DRAM row cycle times and increasing DRAM device datarates, power limitation constraints, and data burst lengths in future generations of DRAM devices. Simulation results obtained in the analysis provide insights into DRAM memory system performance characteristics including, but not limited to the following observations.

The performance benefit of having a 16 banks over 8 banks increases with increasing datarate. The average performance benefit reaches 18% at 1 Gbps for both open-page and close-page systems.

Close-page systems are greatly limited by DRAM device power constraints, while open-page systems are less sensitive to DRAM device power constraints.

Increasing burst lengths of future DRAM devices can adversely impact cache-limited processors despite the increasing bandwidth. Performance losses of greater than 50% are observed.

Finally, This dissertation also present a unique rank hopping DRAM command-scheduling algorithm designed to alleviate the bandwidth constraints in DDR2 and future DDRx SDRAM memory systems. The proposed rank hopping scheduling algorithm schedules DRAM transactions and command sequences to avoid the power limiting constraints and amortizes the rank-to-rank switching overhead. Execution based simulations show that some workloads are able to fully utilize the additional bandwidth and significant performance improvements are observed across a range of workloads.