Software Dsm Behavior Analysis And System Performance Optimization

Software DSM technique provides a method to construct convenient parallel programming environments for low cost computers (workstations or PCs) and networks (LANs). However, the performance of software DSM systems is unsatisfied due to the high overhead of remote memory access. This dissertation investigates the behavior of shared data in software DSM systems and proposes methods to improve the performance of software DSM systems through examining consistency protocols, algorithms of remote access, and applications. The main conclusion of this research is that the performance of software DSM systems can be significantly improved when techniques from all aspects are combined. However, the scope of suitable applications is still confined to regular and coarse-grained remote access in software DSM systems connected by a low speed LAN. The contributions achieved in this dissertation are as follows.First, the dissertation describes the software overheads in time and space, proposes a method to analyze the overheads of remote access in release consistency (RC) and lazy release consistency (LRC), and sets up a test-bed based on CVM, a well accepted software DSM, to examine the result of quantitative analysis and the effect of performance-enhancing techniques.Second, a new multiple-writer approach, called Limited Multiple-Writer (LMW), is proposed to alleviate the impact of false sharing. LMW distinguishes lock-based and barrier-based false sharing, and handles them with different policies: RC and LRC, single-writer and multiple-writer. Meanwhile, most Twin and Diff employed in regular multiple-writer approach are discarded and the implementation of multiple-writer is simplified.Third, a method of pre-processing shared data, called pre-sending approach, is presented. It is based on the analysis of shared data behavior and the labeling of shared array. In this way, the characteristics of shared variables in high level of a system, such as HLL, can be dynamically passed into the lower level (DSM), which enhances the locality of the shared data and reduces the number of remote accesses. Two kinds of prototype are implemented on HP workstations and SPARCstation-10s.