Performance measurement and analysis of a shared-memory multiprocessor

The performance analysis of the VPS (Virtual Processor System) multiprocessor is presented. The system incorporates four to twenty 8-MHz 8086/8087 processor boards and is built with two main objectives in mind: speed and transparent operation. The von Neumann shared bus bottleneck is alleviated by executing the code in private memories and distributing the synchronization using dual-port memories. Furthermore the system is easily programmable using the LEM analyzer, an automatic program partitioner especially developed for the VPS. Both hardware and software performance measures were taken, using detailed minimal-intrusive timing probes and emulation traces. The test covered six programs, including linear system solving, sorting, matrix multiplication and Fast Fourier transformation. Apart from the raw speedup figures, four sources of overhead were analyzed: the processor idle time, the synchronization overhead, the bus delay and the task-switch time. The results show that the task granularity has the largest impact on the performance