DeepSwapper: A deep learning based page swap management scheme for hybrid memory systems

Abstract

In this paper, we introduce DeepSwapper, a deep learning-basedpage swap management scheme that utilizes RNN to performfast, energy-efficient, and temperature-aware page swapping inhybrid memory systems. DeepSwapper comprises of LSTM unitsof RNN model to predict the future memory accesses to guide itsswap management scheme, a dynamic page swap managementscheme that utilizes DRAM capacity efficiently by enabling hotpages in a swap group to be swapped with cold pages of anotherswap group, and a temperature-aware page swap managementscheme, which first predicts the future writes to NVM pages andthen, decides to migrate those pages with frequent writes in hotNVM banks to DRAM to enhance the NVM lifetime.