Tuning NUMA Memory Management for Applications and Architectures

\indent The study of operating systems level memory management policies for nonuniform memory access time (NUMA) shared memory multiprocessors is an area of active research. Previous results have shown that the best policy choice often depends on the application under consideration, while others have reported that the best policy depends on the particular architecture. Since both observations have merit, we explore the concept of policy tuning on an application/architecture basis. We introduce a highly tunable dynamic page placement policy for NUMA multiprocessors, and address issues related to the tuning of that policy to different architectures and applications. Experimental data acquired from our DUnX operating system running on two different NUMA multiprocessors are used to evaluate the usefulness, importance, and ease of policy tuning. Our results show that both application and architectural dependencies arise and interact with each other. Tuning policies for specific applications is justified by the benefits that have been realized and by the differences observed in how the programs of our workload respond to different parameter settings. The need to tune policies for the architecture seems obvious and is supported by data comparing how differences in costs affect the ``right'''' management decisions for the same application running on each architecture. We believe that a parameterized policy, such as that in DUnX, can form the basis for machine-independent memory management for NUMA machines.