Work as a team or individual: Characterizing the system-level impacts of main memory partitioning

Abstract

Modern multi-core systems employ shared memory architecture, entailing problems related to the main memory such as row-buffer conflicts, time-varying hot-spots across memory channels, and superfluous switches between reads and writes originating from different cores. There have been proposals to solve these problems by partitioning main memory across banks and/or channels such that a DRAM bank is dedicated to a single core, being free from inter-thread row-buffer conflicts. However, those studies either focused on only multi-programmed workloads on which cores operate independently, not cooperatively, or specific hardware configurations with a limited number of degrees of freedom in the number of main memory banks, ranks, and channels. We analyze the influence of memory partitioning on systems with various degrees of banks, ranks, and channels using multi-threaded and multi-programmed workloads, making the following key observations. Bank partitioning is beneficial when memory-intensive applications in a multi-programmed workload have similar characteristics in bank-level parallelism, bandwidth, and capacity demands. Any diversity in these demands with a limited memory capacity greatly diminishes the bank partitioning benefits. As memory access/usage patterns across cores are more easily manageable on multi-threaded workloads, bank partitioning is more often effective with memory intensive multithreaded applications. Channel partitioning becomes effective when the reduction of the negative impacts of time-varying hotspots across memory channels outweighs the load imbalance due to partitioning. We also demonstrate the benefits of rank partitioning with regard to minimizing read-write switches on multi-threaded applications where cores can coordinate memory accesses.