Cooperative NV-NUMA: prolonging non-volatile memory lifetime through bandwidth sharing

Resistive memory technologies, such as ReRAM and PCM, are potentially promising replacements for DRAM technology. Their limited endurance (and thus short lifetime), however, is a major obstacle to their commercialization. Analytic models and experimental data show a polynomial relationship between write latency and endurance. Thus, we can use slow writes to introduce less wear, but it is challenging to design a system that meets the memory lifetime requirements without losing performance. We address this challenge in a multiprocessor non-uniform memory architecture (NUMA) environment through memory bandwidth sharing between processing nodes.

While previous approaches have distributed data and computation to share memory capacity or parallelize applications, our main goal is to share memory bandwidth between NUMA nodes when running workloads with varying memory bandwidth needs. When a node has extra memory bandwidth and a write is for data residing in that node, then a slow write can be issued and lifetime can be improved. Data distribution, however, creates a new challenge in that latency for remote nodes' memory accesses can degrade performance. In order to mitigate this degradation, we propose Cooperative NV-NUMA, which detects hot remote memory pages by monitoring Last Level Cache (LLC) evictions, and caches these pages locally. We simulate a proof-of-concept design that explores the proposed technique for a suite of applications. We find that, our approach can meet lifetime requirements while substantially improving the performance of the NUMA nodes with challenging lifetime (up to 48%) over previous work.