权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Prototype Development and Experimental Implementation of A Massively Parallel Computer

大规模并行计算机的原型开发和实验实现

基本信息

批准号：
06508001
负责人：
TOMITA Shinji
金额：
$ 51.65万
依托单位：
KYOTO UNIVERSITY
依托单位国家：
日本
项目类别：
Grant-in-Aid for Developmental Scientific Research (A)
财政年份：
1994
资助国家：
日本
起止时间：
1994 至 1995
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-06508001/
关键词：
Massively Parallel Computer Workstation Cluster Distribute Memory System Cache Memory Distributed Simulation Maintenance System Scalability I O Subsystem スケーラビィリティキャッシュ Rocursive Dragonal Torus ルーティングマルチキャストメンテナンス統計データの可視

项目摘要

The research result in this year is as follows :1. Development environment for massively parallel system software We have constructed a distributed shared memory environment on an usual workstations with an interconnection network for massively parallel computers named Recursive Diagonal Torus (RDT) and a super high-spedd I/O system named STAFF-link, which are the fruits of the grant-in-aid for scientific research on priority areas from 1992 to 1994 and this grant-in-aid for developmental scientific research, to develop hardware subsystem which enales software development for massively parallel computer systems.2. Improvement of pseudo full-map directory coherence control We have proposed the pruning cache scheme which reduces unneeded multicast message of pseudo full-map (hierarchical map) scheme suitable for RDT,and verified its efficiency.3. Experimental implementation of a large scale file system We have developped a super high-speed disk subsystem which drives 64 disks in parallel with 16 workstations with 4 disk subsystems connected with STAFF-link. This system hides communication latency depending on the inter-node distance by the virtual bi-directional FIFO of STAFF-link, and provides peak transmisson speed of 140Mbps.4. High-density implementation for massively parallel computers We have examined an implementation scheme to consruct the system of CPU boards, cluster boards, and mother boards to realize 1024-processor configuration, and verified its feasibility and confirmed its behavior by implementing real boards.

本年度的研究成果如下：1.大规模并行系统软件的开发环境我们在具有大规模并行机互连网络的普通工作站上构建了一个分布式共享内存环境和一个名为Staff-link的超高速I/O系统，它们是1992-1994年度重点领域科研资助的成果，为大规模并行计算机系统的软件开发提供了硬件子系统。伪全图目录一致性控制的改进我们提出了一种剪枝缓存方案，减少了适用于RDT的伪全图(Hierarchical MAP)方案中不需要的组播消息，并验证了其有效性。一个大型文件系统的实验实现我们开发了一个超高速磁盘子系统，它与16个工作站并行驱动磁盘，有4个磁盘子系统通过员工链接连接。该系统通过员工链路的虚拟双向FIFO隐藏了依赖于节点间距离的通信延迟，并提供了140Mbps的峰值传输速度。大规模并行计算机的高密度实现我们研究了一种构建CPU板、集群板和主板系统来实现1024处理器配置的实现方案，并通过实现真实的板来验证其可行性和行为。