Tornado: Fine-Grain Distributed Shared Memory for SMP Clusters

点击翻译按钮获取中文摘要

This project is exploring the cluster approach to symmetric multiprocessing, in which bus-based multiprocessors are interconnected using a more scalable network. Using the Tempest parallel programming substrate, the project is building clusters that extend the shared memory model with fast references to recently used locations across a cluster. The key issue is achieving good performance without resorting to extensive custom hardware. The project encompasses design and some implementation of a high-speed Tempest system that can: provide shared memory across a cluster; use cacheable control registers and queues for communication among processors and I/O devices; and provide hardware acceleration of protocols. The project includes the development of programming tools and compilers that facilitate specification and verification of custom protocols, and use these protocols to hide latency.