Bandwidth-aware DRAM page migration for heterogeneous mobile memory systems

Abstract

Bandwidth demands for mobile consumer electronics devices, including smartphones and tablets, have been continuously increasing in recent years. A heterogeneous memory system (HMS) integrating different types of mobile DRAM devices, such as LPDDR4 and WideIO2, is an attractive solution to scale DRAM bandwidth beyond the maximum bandwidth of a single device type. One of the key challenges in HMS design is how to identify hot (cold) pages and migrate them to the faster (slower) devices at a low cost. This becomes more difficult as more devices like LPDDR4X adopt dynamic voltage and frequency scaling (DVFS) for DRAM I/O to reduce power consumption. At a low I/O voltage and frequency, a significant portion of DRAM bandwidth can be consumed by page migration traffic to slow down user applications. Thus, this paper proposes a novel page migration technique towards bandwidth-aware memory management (BAMM) for an HMS composed of both fast WideIO2 and slow LPDDR4 devices. BAMM periodically measures the bandwidth consumption for the LPDDR4 channel and adjusts both the maximum cap for migration bandwidth and DVFS level to keep the migration overhead low. Our evaluation using a detailed cycle-level simulator demonstrates that BAMM improves the energy efficiency of the memory system by 6.3% on average (and up to 11.6%) with minimal performance overhead over the baseline HMS with no dynamic page migration.