Hierarchical collections: an efficient scheme to build an object-oriented distributed class library for massively parallel computation

摘要

Separation of parallelism and distribution is one of major concerns of efficient massively parallel computation. The details of distribution should be hidden from users of parallel/distributed class frameworks, while they should be easily modifiable by (library) programmers who are builders of the framework. We propose a new scheme to build an object-oriented parallel distributed class framework based on a simple but mathematically disciplined model called hierarchy of collections. Based on the model, classes can be easily derived to achieve high performance massively parallel computation on a variety of physical platforms. We have examined the descriptive power of our proposal with various specialized distributions including the recently proposed Twisted Data Layout, on the Fujitsu AP1000 parallel computer.