Performance evaluation of home-cluster based scheduling for NUMA multiprocessors

For Non-Uniform Memory Access (NUMA) multiprocessors, memory access overhead is crucial to system performance. The memory access overhead strongly depends on the processor scheduling scheme and the page placement scheme. When a thread in a process accesses virtual pages allocated in remote memories during its execution, the thread is not efficiently executed due to remote memory access overhead. Therefore, the processor scheduling scheme and the page placement scheme should work cooperatively for efficient process execution. A home-cluster based processor scheduling is a kind of cooperative scheduling scheme that works with the page placement scheme using the notion of the home-cluster. This paper investigates the performance of the home-cluster based policy that consists of a processor scheduling scheme and a page placement scheme that are based on the home-cluster for cluster-based NUMA multiprocessors. The home-cluster based policy is compared with other existing policies, which focus on only processor scheduling rather than page placement scheme. Those policies are examined through the simulation experiments by using a process execution model, MVA process model). The simulation results show that mean response time with the home-cluster based policy is shorter than that with the other existing ones because of reducing remote memory access overhead. That means that the home-cluster based policy gives better system performance than the other existing ones.