Reinforcement Learning Based Data Compression for Energy-Efficient Non-volatile Caches.
Fanfan Shen,Chao Xu,Jun Zhang,Yong Chen,Yanxiang He
DOI: https://doi.org/10.1109/hpcc-dss-smartcity-dependsys53884.2021.00110
2021-01-01
Abstract:Non-volatile cache (NVC) technologies are promising candidates to build the next generation cache architecture due to its high density, non-volatile and low leakage power consumption. However, NVC consumes high write latency and energy. Existing data compression schemes are an efficient way to solve the NVC's problem, for reducing bit-write and expanding cache capacity. Nevertheless, they still suffer from limited compression performance with low compression ratio. In order to overcome these shortcomings, we propose a Reinforcement Learning based Data Compression scheme (RLDC), including data collection, reinforcement learning model, prediction table, compression and decompression logic. The idea behind RLDC is to explore the characteristics of write intensive words and obtain data compression patterns in the execution flow. Evaluation results show that the proposed scheme achieves compression ratio by 1.88X, reduces energy consumption by 37.57% and execution time by 37.01% compared with the baseline.