Performance study of a concurrent multithreaded processor

The performance of a concurrent multithreaded architectural model, called superthreading, is studied in this paper. It tries to integrate optimizing compilation techniques and run-time hardware support to exploit both thread-level and instruction-level parallelism, as opposed to exploiting only instruction-level parallelism in existing superscalars. The superthreaded architecture uses a thread pipelining execution model to enhance the overlapping between threads, and to facilitate data dependence enforcement between threads through compiler-directed, hardware-supported, thread-level control speculation and run-time data dependence checking. We also evaluate the performance of the superthreaded processor through a detailed trace-driven simulator. Our results show that the superthreaded execution model can obtain good performance by exploiting both thread-level and instruction-level parallelism in programs. We also study the design parameters of its main system components, such as the size of the memory buffer, the bandwidth requirement of the communication links between thread processing units, and the bandwidth requirement of the shared data cache.