User-level Networking For Software DSMs

Clusters are scalable platforms for parallel applications with high performance/price ratio. They are always composed of nodes of workstations that are connected by high performance networks. User-level networking can improve the communication performance of clusters significantly by reducing software overheads and memory copies, from which the high level parallel applications will benefit very well.The software distributed shared memory (DSM) systems on clusters are competitive parallel computing environments that are easy to use and can be applied on most computer platforms. Although many sophisticated schemes have been proposed and the communication pressure between nodes has decreased very much the performance of software DSMs is not yet as good as either hardware DSMs or message-passing systems. User-level networking can improve the performance of software DSMs further that can replace the inefficient communication substrates of UDP/IP.In this dissertation, we study the characteristics of software DSMs thoroughly and then present the implementation and optimization of JIAJIA system. At the same time, the analysis of behavior of JIAJIA system shows how the communication substrates act the important roles in software DSMs. We compare many user-level communication libraries and summarized their key issues. Then we design and implement a user-level communication library called JMCL for JIAJIA system and evaluate its impacts on JIAJIA system.The analysis of JMCL shows that user-level networking significantly improve the communication performance of software DSMs so that the high level applications speed up much more than before, especially the applications that are thought to be unsuitable for software DSMs because of their high pressure of communication. User-level networking shifts the boundary of applications between software and hardware DSMs and make it possible that software DSMs would be convenient and cheap platforms for most parallel applications. Otherwise user-level networking does not change the computing nature of software DSMs. It make software DSMs run faster just by cutting the elapsed time of all parts of software DSMs proportionally through reducing the communication overhead and memory copies. To improve the performance of software DSMs further we have to still focus on the progress of memory consistency models and cache coherence protocols.