Shared-Memory Multiprocessors with Stable Performance

The shared-memory programming model is attractive to programmers of parallel computers because they are not required to control the placement and communication of application data. Unfortunately, access to data is the root cause of performance problems on distributed shared-memory multiprocessors. Severe performance degradation can occur when there are many access requests competing for service at one or a few processing nodes. The request messages must join a queue for service, thereby increasing the delay between the original issue of a remote access request and the receipt of the data. This thesis presents a new architectural mechanism for remote read requests which reduces excessive queueing. The reduction in queueing comes from using intermediate nodes, called proxies, to service read requests. By combining read requests for data at a proxy, only one request has to be forwarded to the data's home node. The use of proxies introduces overheads, primarily due to the extra messages which can arise from sending read requests via a proxy, and the local storage space which a proxy uses to retain data to satisfy later requests from its clients. The overheads can cause performance degradation if proxies are used when they are not needed. To avoid such adverse e ects, several variants of the proxy technique are investigated: invoking proxies automatically when excessive queueing occurs at run-time (reactive and adaptive proxies), not holding proxy data copies, and using separate storage bu ers for proxy data copies. The di erent implementations of the proxy technique are evaluated using execution-driven simulations of eight benchmark applications. The results show that substantial performance bene ts can be obtained by using proxies for applications which have widely-shared data structures, and using adaptive proxies in conjunction with separate proxy bu ers can improve the performance for all the benchmark applications. 3