Efficient management of wormhole routing network in multicomputers

Abstract

Wormhole routing networks have been widely used in many advanced distributed-memory multicomputers and distributed shared-memory multicomputers due to its various advantageous features and excellent performance. To derive an optimal wormhole router model, the performance evaluation of various aspects of wormhole routing networks have to be conducted. For the purpose, a wormhole router simulator is developed using the DEVS (discrete event specification) methodology, that specifies wormhole routing algorithms as well as hardware designs of various features in a formal, hierarchical, and modular manner. Next, a new wormhole router model, which has the ability to multicast by hardware supporting, is proposed. The proposed wormhole router model guarantees multicast deadlock- and starvation-freeness, which cannot be satisfied by other conventional router models. Along with a low-latency and high-throughput hardware router, efficient management of wormhole routing networks is also a critical problem to build a high performance multicomputer. First, the mapping problem is considered under the wormhole routing environment. A new objective function for mapping is derived from the extensive analysis of the blocking property and channel contention, which are unique characteristics of wormhole routing network. Our wormhole router simulator is used to evaluate the performance of mapping results and the characteristics of wormhole routing network are reflected well by our objective function. Second, a new processor allocation scheme in wormhole routed hypercube multicomputers is proposed. The scheme reflects the virtual cut-through capability of wormhole routing, and our algorithm finds additional resources, called modified subcubes, for processor allocation. Simulation results show that, using the scheme, higher processor utilization and better overall performance of a multicomputer system can be achieved. Last part of the thesis is devoted to the network partitioning proble...