Reconfiguration and Performance Issues in Software Distributed Shared Memoroy Systems

软件分布式共享存储器(DSM)可以支持从商品或定制微处理器和网络构建的各种现有的工作站和分布式存储器机网络，而无需添加任何硬件。为了将DSM应用于长时间运行的应用程序或高可用性情况，响应处理器故障或资源变化的重新配置是至关重要且不可避免的。从性能的角度来看，基于软件的DSM系统应该尽量减少其一致性流量开销，因为这样的系统对网络上的流量特别敏感。本研究的目的是研究软件DSM在发布一致性下的重构和性能问题，旨在使这种DSM成为一种有利的计算范例。本研究的主要目的包括：(1)寻求高效的重构算法来识别分布式系统中的目标结构；(2)研究基于多条不相交路径的进程迁移；(3)开发一个基于快速检查点的可恢复的软件DSM系统；(4)研究软件DSM的发布一致性的积极实现，(5)寻求有效的软件需求侧管理的屏障和锁定技术。

Software distributed shared memory (DSM) may support various existing networks of workstations and distributed memory machines built from commodity or customized microprocessors and networks, without any hardware added. In order for DSM to be employed in long-running applications or high-availability situations, reconfiguration in response to processor failures or resource changes is crucial and inevitable. From the performance standpoint, a software-based DSM system ought to minimize its coherence traffic overhead, because such a system is especially sensitive to the traffic amount over the network.This research proposal focuses on reconfiguration and performance issues in software DSM under release consistency, aiming to render such DSM a favorable computing paradigm.The key objectives of this research include: (1) pursuing efficient algorithms for identifying target structures in distributed systems under reconfiguration, (2) investigating process migration by taking advantage of multiple disjoint paths, (3) developing a recoverable software DSM system based on fast checkpointing, (4) investigating into aggressive implementations of release consistency for software DSM, and (5) pursuing efficient barrier and lock techniques for software DSM.