Multi-Level Memory Policies: What You Add Is More Important Than What You Take Out

Abstract

Multi-Level Memory (MLM) will be an increasingly common organization for main memory. Hybrid main memories that combine conventional DDR and "fast" memory will allow higher peak bandwidth at an attainable cost. However, the chief hurdle for MLM systems is the management of data placement. While user-directed placement may work for some applications, it imposes a heavy burden on the programmer. To avoid this burden while still benefiting from MLM, we propose a number of automated management policies. Our results show that several possible policies offer performance and implementation trade offs. Also, unlike conventional cache or paged memory policies, the addition policy is much more important than the replacement policy.