CISE Postdoctoral Program: Scalability of Software Distributed Shared Memory

点击翻译按钮获取中文摘要

9626318 Zwaenepoel, Willy William Marsh Rice University CISE Postdoctoral Program: Scalability of Software Distributed Shared Memory This award supports CES associate Povl Koch. Software Distributed Shared Memory (DSM) architectures strike a balance between the distributed-memory and shared memory architectures by providing an abstraction of shared memory in software on distributed memories. They are built using standard processors/workstations and they are easy to program using a shared address space based on virtual memory pages. DSM systems are suitable for many parallel scientific applications. Although many systems have been designed, only the scalability issues for hardware shared-memory systems have been addressed. For software DSM systems, the focus has been on performance with a moderate number of nodes using relaxed memory consistency models and multiple-writers protocols. Most experiments have been on 8 nodes and a few with 16 nodes. When going from 8 to 64 or more nodes, each may have multiple processors, scalability problems exist with the current size of data structures and their allocation and deallocation. A scalable DSM should only capture actual causality of modifications, minimize space overhead per shared page, and the address space should be dynamically partitioned. We plan to analyze and build a scalable DSM system to run on a distributed-memory machine with 64 nodes and a fast interconnect. ***