Performance of multi-threaded execution in a shared-memory multiprocessor

The authors propose a general analytic model for the overlapped execution of thread management operations in shared-memory multiprocessors. Performance measures as speedup, utilization and network contention are evaluated and verified using simulations. It is assumed that the execution of each thread consists of sequence of single-cycle instructions, interleaved with four high-level thread management instructions: create, activate, suspend and halt (CASH). Each processor (also called the CASH processor) in the multiprocessor is specialized for the efficient execution of the CASH thread execution model. This is achieved by overlapping the execution of all thread management operations in a separate thread management unit, that executes concurrently with the thread execution unit. The execution time of the thread management operations is reduced by increasing the thread memory bandwidth, and by storing multiple thread contexts in a thread cache located on the processor