Automating heterogeneous memory management

Hardware heterogeneity is becoming an increasingly common feature in high-performance computing systems. Unfortunately, while these systems may offer new technologies in memory and computation, the general trend of memory performance is falling behind. With the added complexities of heterogeneous systems, achieving good memory performance is now becoming more difficult. Since different memory technologies exhibit various performance characteristics, careful memory management is required to consider tradeoffs in latency, bandwidth, capacity, and power. Application behavior, including data access patterns, data sizes, and operation types can indicate which characteristics limit the performance of an operation. However, understanding this information and using it to perform optimizations can be a difficult task. We expect this problem to become increasingly prevelent as the memory stack continues to change and expand. In this dissertation we present a solution to managing memory for these heterogeneous systems in an automated manner. We demonstrate its use on several applications and machine types, showcasing the benefit, flexibility, and expandability of the framework.U of I OnlyAuthor requested U of Illinois access only (OA after 2yrs) in Vireo ETD syste