On measuring and evaluating synchronization and virtual memory performance of a multiprocessor with multistage interconnection network

Abstract

Two important system effects on a multiprocessor with multistage interconnection network are studied and examined: the critical section and barrier for synchronization, and processor locality and virtual memory computing for memory management. Several synchronization algorithms using the interconnection network have been evaluated and implemented. The performance of these algorithms has been measured and compared through several numerical examples. Parallel computing makes a new type of locality, called processor locality which requires that data references to memory from a single processor. The issue of processor locality on the interconnection network has been investigated experimentally. The virtual memory computing effects are measured and evaluated on the multiprocessor system. All the experiments and measurements have been done on the BBN GP1000, a shared-memory multiprocessor with multistage interconnection network.<>