Running Parallel Applications on an Mp with Multithreaded Superscalar Processors Running Parallel Applications on a Mp with Multithreaded Superscalar Processors

With lesser returns on adding more complexity to conventional superscalar processors, simultaneous multithreaded (SMT) superscalar processors seem to be a promising alternative. Unfortunately, most previous work has focused on systems running multiprogrammed loads of sequential applications. It is not clear how well these processors work in a shared-memory multiprocessor environment running parallel applications. This paper examines this issue for automatically parallelized and explicitly parallel applications. This paper shows that SMT-based multiprocessors deliver higher performance than multiprocessors with conventional superscalars and are more robust in the presence of low-performing memory systems. Furthermore, while chips with multiple processors called superchips are simpler than chips with an SMT processor, the performance of a superchip-based system running memory intensive parallel applications is much lower than that of a plain SMT-based system. In addition, superchip-based systems are not a good option for workloads with sequential applications. Finally, we propose hardware support to improve the synchronization performance of SMT processors. With this optimization, SMT-based multiprocessors run an average of 14% faster than superchip-based multiprocessors. Overall, SMT-based multiprocessors are the most cost-eeective organization to run parallel applications. Abstract With lesser returns on adding more complexity to conventional superscalar processors, simultaneous multithreaded (SMT) superscalar processors seem to be a promising alternative. Unfortunately , most previous work has focused on systems running multiprogrammed loads of sequential applications. It is not clear how well these processors work in a shared-memory multiprocessor environment running parallel applications. This paper examines this issue for automatically parallelized and explicitly parallel applications. This paper shows that SMT-based multiprocessors deliver higher performance than multi-processors with conventional superscalars and are more robust in the presence of low-performing memory systems. Furthermore, while chips with multiple processors called superchips are simpler than chips with an SMT processor, the performance of a superchip-based system running memory intensive parallel applications is much lower than that of a plain SMT-based system. In addition, superchip-based systems are not a good option for workloads with sequential applications. Finally, we propose hardware support to improve the synchronization performance of SMT processors. With this optimization, SMT-based multiprocessors run an average of 14% faster than superchip-based multiprocessors. Overall, SMT-based multiprocessors are the most cost-eeective organization to run parallel applications.