Memory cocktail therapy: a general learning-based framework to optimize dynamic tradeoffs in NVMs

Non-volatile memories (NVMs) have attracted significant interest recently due to their high-density, low static power, and persistence. There are, however, several challenges associated with building practical systems from NVMs, including limited write endurance and long latencies. Researchers have proposed a variety of architectural techniques which can achieve different tradeoffs between lifetime, performance and energy efficiency; however, no individual technique can satisfy requirements for all applications and different objectives. Hence, we propose Memory Cocktail Therapy (MCT), a general, learning-based framework that adaptively chooses the best techniques for the current application and objectives.

Specifically, MCT performs four procedures to adapt the techniques to various scenarios. First, MCT formulates a high-dimensional configuration space from all different combinations of techniques. Second, MCT selects primary features from the configuration space with lasso regularization. Third, MCT estimates lifetime, performance and energy consumption using lightweight online predictors (eg. quadratic regression and gradient boosting) and a small set of configurations guided by the selected features. Finally, given the estimation of all configurations, MCT selects the optimal configuration based on the user-defined objectives. As a proof of concept, we test MCT's ability to guarantee different lifetime targets and achieve 95% of maximum performance, while minimizing energy consumption. We find that MCT improves performance by 9.24% and reduces energy by 7.95% compared to the best static configuration. Moreover, the performance of MCT is 94.49% of the ideal configuration with only 5.3% more energy consumption.