Communication-Efficient Arbitration Models for Low-Resolution Data Flow Computing.

摘要

Low-resolution data flow computing offers a practical compromise between conventional control-flow computing models and the specialized architectures required for fine-grain data flow processing. We give a formal specification of an arbitration facility that simultaneously partitions and statically assigns operations to processors. This general model is based on differences in the processors, diversity of data links in the network, size of tokens flowing between nodes in the data flow graph, memory limitations on the processors, and considerations to promote parallelism. A network model solves the static problem for bipartite and tree structured data flow graphs. Based on this centralized static allocation scheme, data tokens are automatically routed to processors, and the run-time scheduling process is distributed among the processors. Dynamic arbitration implemented as a centralized facility takes inadequate advantage of network capabilities. A general decentralized (distributed) dynamic arbitration scheme maps tasks to processors at run-time, with the association of tasks to processors based on intertask communication and network data link characteristics. Task migration is supported by treating both data and code as tokens. No centralized control or mass storage are required in this communication-efficient arbitration model. (KR)