Skip to main content
U.S. Department of Energy
Office of Scientific and Technical Information

Linked-crossbar architecture for multicomputer interconnection

Thesis/Dissertation ·
OSTI ID:5826502
A major problem in designing large-scale parallel systems is the construction of an interconnection network to provide interprocessor communications and, in some cases, memory access for the processors. Crossbar switches have been successfully used in systems with a small number of Processing Elements (PEs) but have been ruled out as a means for connecting large numbers of PEs due to their complexity and cost. Large Interconnection Networks (INs) using such connection topologies as n-cube and mesh are more economical but are relatively slow. A Linked Crossbar Architecture (LCA) is proposed in this dissertation to alleviate this large IN problem. LCA consists of many smaller crossbar switches, each connecting a cluster of PEs. The smaller crossbars in LCA can be single chip VLSI components, whose size and capability can move with technology. While the advantage of a crossbar network is apparent at the local level, the goal is to turn this local advantage into a global communication advantage. To achieve this goal, three issues are identified: the communication efficiency, the suitable topologies, and the routing flexibility. A statistical model is developed to analyze the communication efficiency of packet switching for LCA. A new LCA topology is proposed and analyzed, which shows an order of magnitude reduction of message delay when compared to link-oriented INs with similar topologies. This research demonstrates that the Linked Crossbar Architecture is a viable method to build high performance INs relying on the powerful and still growing VLSI technology for future multicomputers. An equally important conclusion is that, with LCA, large scale INs can be built on the low cost and high performance VLSI crossbar switches.
Research Organization:
Southern Methodist Univ., Dallas, TX (USA)
OSTI ID:
5826502
Country of Publication:
United States
Language:
English