Realization of synchronization tools and their efficiency in the multiprocessor system M5Ps
Johannes Milde, Lothar Krings, Walter Ameling · 1982
One reason for development of parallel processor systems is to increase the performance or the ratio between performance and cost to overcome the limits of a sequential single processor system. This is possible because many important applications (matrix manipulation, image processing, process control) can be solved by a set of widely independent processes. Normally these processes must be synchronized at some points in order to exchange information or to guarantee a consistent computation. This synchronization and communication reduce the possible speed-up of a parallel algorithm execution. Therefore it is necessary to implement efficient synchronization tools in a multiprocessor system. In the multiprocessor system M5Ps the solution of the synchronization problem is organized in several layers and realized by hardware and software. In the MIMD-operating mode it can be achieved by the two abstract data types lock (realized by hardware), and semaphore (realized by the operating system). In the SIMD-operating mode all processors are synchronized before each instruction. After a data dependent branch, two sets of processors can operate in SIMD- and MIMD-mode simultaneously, and join later again. 4 references.