Memory Management in Hybrid DRAM/NVM Systems using LSTMs

Abstract

Heterogeneous memory technologies have been widely used in effort to address the ever-increasing demands of modern applications for larger main memory capacity. The new technologies showcase vastly greater differences in terms of capacity, latencies and bandwidth. This heterogeneity along with the the greater irregularity of emerging workloads, render state-of-the-art memory management solutions insufficient; thus calling for more intelligent methods. In this diploma Thesis, we design and evaluate a scheduler which intelligently places application data, on a Page granularity, across hybrid memory components using Artificial Neural Networks. The proposed Scheduler combines intelligent page placement decisions leveraging LSTM networks with existing history-based data tiering methods. The scheduler focuses the machine learning on a page subset whose timely movement will reveal most application performance improvement, while being mindful of computation resources. K-Means address space clustering is also utilized to augment the eviction policy used by the proposed scheduler in order to provide application performance boost. That boost is on average 10% according to our evaluation process. Our performance evaluation also indicates that the proposed Scheduler significantly reduces the performance gap between existing solutions and an oracle scheduler with a priori knowledge of the page access patterns, while being a potential candidate for designing low-power oriented Hybrid Memory Systems as well.