PYI: Automatic Locality Management in Scaleable Multiprocessors

点击翻译按钮获取中文摘要

Agarwal Parallel computers can be made both scalable and easily programmable through architectures that exploit and automatically manage communication locality. The goal of this research is to discover and to evaluate techniques for automatic locality management in scalable multiprocessors. As the vehicle for this research, an experimental parallel machine called the Alewife is being implemented. Alewife employs techniques for: 1. communication latency minimization, using scalable coherent caches and software partitioning and placement of programs, and 2. communication latency tolerance, using a new rapid-context- switching processor architecture. Alewife implements a new protocol called "limitless directories" for scalable cache coherence. This scheme uses a combination of hardware and software techniques to realize the performance of a full-map directory with the memory overhead of a limited directory. A rapid-context-switching processor called Sparcle is also being designed. Sparcle can switch in about 10 cycles to another thread when it suffers a cache miss that requires service over the interconnection network. The major goal for this grant period is to get a small prototype Alewife system operational, including the hardware as well as the entire software system.