Supporting Fine-Grained Synchronization on a Simultaneous Multithreading Processor

Existing multiprocessor synchronization mechanisms are relatively heavyweight, due in part to the level of the memory hierarchy (typically main memory) at which threads must synchronize. Multithreaded processors, on the other hand, have the potential to significantly reduce synchronization cost, because threads share the processor simultaneously and can synchronize using processor-internal state. This paper proposes and evaluates new synchronization schemes for a simultaneous multithreaded processor. We present a scalable mechanism that permits threads to cheaply synchronize within the processor, with blocked threads consuming no processor resources. The basic mechanism, blocking acquire and release, is a hardware implementation of traditional software synchronization abstractions, implemented with a thread-shared hardware lock box. This study shows that a multithreading processor with efficient synchronization enables parallelization at a granularity an order of magnitude finer than required on conventional parallel machines with memory-based synchronization. When combined with lock release prediction, which reduces the overhead of restarting a blocked thread, acquire-release gains an additional improvement of 40%. Overall, we show that these substantial improvements in synchronization cost enable parallelization of code that could not be effectively parallelized using traditional synchronization; performance of these codes showed up to a two-fold improvement over single-threaded versions of the previously unparallelizable code.