Modeling memory system performance of NUMA multicore-multiprocessors

The performance of many applications depends closely on the way they interact with the computer’s memory system: Many applications obtain good performance only if they utilize the memory system efficiently. Unfortunately, obtaining good memory system performance is often difficult, as developing memory system-aware (system) software requires a thorough and detailed understanding of both the characteristics of the memory system and of the interaction of applications with the memory system. Moreover, the design of memory systems evolves as newer processor generations appear on the market, thus the problem of software–hardware interaction must be revisited to understand the interaction of (already existing) software with newer memory system designs as well. This thesis investigates the memory system performance of a recent class of machines, multicore-multiprocessors with a non-uniform memory architecture (NUMA). A NUMA multicore-multiprocessor system consists of several processors where each processor integrates multiple cores. Typically, cores of a multicore processor share resources (e.g., last-level caches) and contention for these shared resources can result in significant performance degradations. NUMA multicore-multiprocessors are shared-memory computers, but the memory space of a NUMA multicore-multiprocessor system is partitioned between processors. Accessing the memory of a local processor takes less time than accessing the memory of other (remote) processors, therefore data locality (a low number of remote memory accesses) is critical for good performance on NUMA machines. This thesis presents a performance-oriented model for NUMA multicore-multiprocessors. The model considers two application classes, multiprogrammed workloads (workloads that consist of multiple, independent processes) and multithreaded programs (programs that consist of a number of threads that operate in a shared address space). The thesis presents an experimental analysis of memory system bottlenecks experienced by each application class. Moreover, the thesis presents techniques to reduce the performance-degrading effects of these bottlenecks. We determine (based on experimental analysis) that the performance of multiprogrammed workloads depends on both multicore-specific and NUMA-specific aspects of a NUMA multicore-multiprocessor’s memory system. Therefore, a process scheduler must find a balance between reducing cache contention and improving data locality; the N-MASS scheduler presented by the thesis attempts to strike a balance between these, sometimes contradicting, goals. N-MASS improves performance up to 32% over the default setup in current Linux implementations on a recent 2-processor 8-core machine. Based also on experimental analysis we find that data locality is of prime importance for the performance of multithreaded programs. The thesis presents extensions to two popular parallel programming frameworks, OpenMP and Intel’s Threading Building Blocks. The extensions