Design trade-offs of low-cost multicomputer network switches

A comparison is made among a large number of designs for the purpose of specifying low-cost yet cost-effective multicomputer network switches. Among the parameters varied are switching mode, number of lanes, buffer size, wraparound, and channel selection. Some assumptions we make are deterministic routing and small fixed-sized packets. We obtain results using two methods: i) RTL cycle-driven simulations to determine latency and capacity with respect to load, communication pattern, and packet size and ii) hardware synthesis to a current technology to find the operating frequency and chip area. These results are also combined to yield performance/area measures for all of the designs. One of the results is deeper understanding of virtual cut-through in terms of deadlock properties and the capability of dynamic load balancing among buffers. We find that lanes are even more likely to improve performance of virtual cut-through as wormhole networks, and that virtual cut-through routing is preferable to wormhole routing in more domains than may have been previously realized. Other results include finding that, after factoring in operating frequency, having more than two lanes per physical channel is not likely to be useful and various observations about the utility of varying buffer sizes.

[1]  William J. Dally,et al.  Deadlock-Free Message Routing in Multiprocessor Interconnection Networks , 1987, IEEE Transactions on Computers.

[2]  A. A. Chein,et al.  A cost and speed model for k-ary n-cube wormhole routers , 1998 .

[3]  J DallyWilliam,et al.  Performance Analysis of k-ary n-cube Interconnection Networks , 1990 .

[4]  Kang G. Shin,et al.  Tailoring router architectures to performance requirements in cut-through networks , 1996 .

[5]  Lawrence Snyder,et al.  The Case for Chaotic Adaptive Routing , 1997, IEEE Trans. Computers.

[6]  Frank Thomson Leighton,et al.  Average case analysis of greedy routing algorithms on arrays , 1990, SPAA '90.

[7]  Andrew A. Chien,et al.  A Cost and Speed Model for k-ary n-Cube Wormhole Routers , 1998, IEEE Trans. Parallel Distributed Syst..

[8]  Martin C. Herbordt,et al.  An empirical study of datapath, memory hierarchy, and network in SIMD array architectures , 1995, Proceedings of ICCD '95 International Conference on Computer Design. VLSI in Computers and Processors.

[9]  Andrew A. Chien,et al.  The Cost of Adaptivity and Virtual Lanes in aWormhole Router , 1995 .

[10]  Sudhakar Yalamanchili,et al.  Interconnection Networks: An Engineering Approach , 2002 .

[11]  William J. Dally Virtual-Channel Flow Control , 1992, IEEE Trans. Parallel Distributed Syst..

[12]  William J. Dally,et al.  Performance Analysis of k-Ary n-Cube Interconnection Networks , 1987, IEEE Trans. Computers.

[13]  Kevin Bolding Non-Uniformities Introduced by Virtual Channel Deadlock Prevention , 1992 .

[14]  Leonard Kleinrock,et al.  Virtual Cut-Through: A New Computer Communication Switching Technique , 1979, Comput. Networks.

[15]  Steven L. Scott,et al.  Optimized Routing in the Cray T3D , 1994, PCRCW.

[16]  William J. Dally,et al.  The message-driven processor , 1992 .