NCESPARC+: an implementation of a SPARC architecture with hardware support to multithreading for the multiplus multiprocessor

Abstract

NCESP ARC + is an implementation of the SP ARC v: 8 architecture with hardware support to a variable number of thread contexts, which is under development for use within the framework of the Multiplus distributed shared-memory multiprocessor. It is expected to provide an efficient and automatic mechanism to hide the latency of busy-waiting synchronization loops, cachecoherence protocol and remote memory access operations within the Multiplus multiprocessor. NCESPARC + performs context-switching in at most four processor cycles whenever there is an instruction cache miss, a data dependency in relation to the destination operand of a pending load instruction or a busy-waiting synchronization loop. It has a decoupled architecture which allows the main pipeline to process instructions from a given context while the Memory Interface Unit performs memory access operations related to that same context or to any other context. Results of simulation experiments show the impact of some architectural parameters on the NCESPARC + processor performance and demonstrate that the use of multiple thread contexts can e.ffectively produce a much better utilization of the processor when long latency operations are performed In addition, NCESPARC + processor performance with a single context is superior to that of a standard implementation of the SPARC architecture due to its decoupled architecture.