A New Synchronization Scheme for Memory Consistency Model ( Extended Abstract )

Modernistic scalable multiprocessors are mostly built with a distributed-shared memory architecture. Large scale shared memory multiprocessors have long memory latencies for the remote memory access. And these latencies can quickly offset system performance earned from the exploitation of parallelism. In order to improve system performance, we must reduce memory latencies. The useful way for the improvement is to use efficient synchronization scheme. However, the past relaxed memory consistency models are not suitable for the out of order execution environment. In this paper, we present a new synchronization scheme based on superscalar multiprocessors to exploit the parallelism. Several synchronization primitives are proposed to create a more weak memory consistency environment. Extra hardware is added to support the new synchronization operation. A simulator is developed to evaluate the performance enhancement of various synchronization schemes. Experimental results display that the performance improvement of our proposed synchronization scheme is significant.