TOWARD A MORE REALISTIC PERFORMANCE EVALUATION OF INTERCONNECTION NETWORKS

Interconnection network design plays a central role in the design of p arallel systems. Most of the previous research has evaluated the perfo rmance of interconnection networks in isolation. in this study, we inv estigate the relationship between application program characteristics and interconnection network performance using an execution driven simu lation testbed: the Reconfigurable Architecture Workbench (RAW). We si mulate five topological configurations of a k-ary n-cube interconnect and four different network [ink models for a 4,096 node SIMD machine, and quantify the impact of the network on two application programs. We provide experimental evidence that such ''in-context'' simulation pro vides a better view of the impact of network design Variables on syste m performance. We show that recent results, indicating that low-dimens ional designs provide better ICN performance, ignore application requi rements that may favor high-dimensional designs. Furthermore, applicat ions that would appear to favor low-dimensional designs may not, in fa ct, be significantly impacted by the network's dimensionality. We expe rimentally test the results of published performance models comparing the use of a synthetic load to that of a toad generated by a typical a pplication program. The experiments indicate that the standard metric of average message latency can vary considerably under different appli cation loads and that average message latency may not reflect overall application performance. In particular, at the level of the offered (a pplication generated) load to the network, the topological properties of the network are important in determining the average message latenc y. However, for overall application performance, we found that the net work topology may not be critical so long as there is sufficient netwo rk bandwidth. in such cases, the results suggest that optimizing the i mplementation cost of the network should be the key design criterion. We also present a simple abstraction for the network that captures all the important design parameters of the interconnect that can be easil y incorporated into any execution-driven simulation framework.