Software distributed shared memory systems with performance scalability for supercluster systems

针对超级集群系统的实用性，实现了一个具有性能可扩展性的高性能软件分布式共享内存（DSM）系统Mocha作为实验原型。当前版本的Mocha工作在Linux环境下的32节点集群上。源程序已通过web在公共领域开放。提出了一种抑制通信开销的方案，该方案主要是为了保持分布内存中的共享数据之间的数据一致性而引起的。通过一些基准程序对其有效性进行了评价。实现了一种用于应用程序性能调试和构建高效软件DSM的工具S-Cat，该工具具有可视化显示应用程序执行过程中处理器间通信轨迹的功能。S-Cat应用于Mocha的高效实现和应用程序在Mocha上的性能调优。主要研究成果如下：1。实现一个低通信开销的软件DSM，命名为Mocha2。同一集群上不同软件dsm的比较研究基于DSM4软件的粗粒度并行处理方案。构建DSM软件可视化工具S-Cat5。在图形处理器和现成处理器上并行处理的建议明确了实现超集群系统高性能软件决策管理的一些技术问题。提出了控制多集群的“分层范式”，但其有效性尚不能在研究阶段得到验证。

A high-performance software distributed shared memory (DSM) system with performance scalability, named Mocha, was implemented as an experimental prototype aiming at the usefulness of supercluster systems. The current version of Mocha works on a 32 node cluster on the Linux environment. The source program has been opened in public domain via web. A scheme for suppressing communication overhead, which is mainly caused by maintaining data consistency among shared data located at distributed memories, is proposed and slotted in Mocha. The effectiveness is evaluated by using some benchmark programs. A tool for performance debugging of application programs and constructing an efficient software DSM, named S-Cat, was implemented, which has such functions as showing visually communication locus among processors during execution process of an application program. S-Cat was used in efficient implementation of Mocha and performance tuning of application programs on Mocha.The main research results are as follows:1.Implementation of a low communication overhead software DSM, named Mocha2.Comparison study of different software DSMs on the same cluster3.Proposal of coarse-grain parallel processing on software DSM4.Construction of a visualization tool for software DSM, named S-Cat5.Proposal of parallel processing on graphic processor and off-the-shelf processorSome technical issues toward realization of high-performance software DSMs for supercluster systems are made clear. "Hierarchical paradigm" for controlling multiple clusters was proposed, but its usefulness cannot be verified during this research period.