PROBLEMS WITH COMPARING INTERCONNECTION NETWORKS - IS AN ALLIGATOR BETTER THAN AN ARMADILLO

Designing a parallel machine would be much easier if one interconnecti on network were ''best'' for all applications and all operating enviro nments (including hardware, software, and financial factors). Unfortun ately, no such network exists. Furthermore, even for a fixed applicati on domain and a fixed operating environment, selecting the best networ k may be difficult because many cost and performance metrics could be used. Suppose someone asked you to select the best animal. What featur es would you use to compare, say, an alligator and an armadillo? Zn so me ways, the two are very similar: both have four legs, a rugged exter ior, and sharp claws. However, in other ways the two are very differen t: one prefers a marshy environment, the other dry land; one has a lon g tail, the other a short tail; one is a reptile, the other a mammal. Which of the two, then, is a better animal and what makes it better? T hese questions apply to interconnection networks as well. Suppose you are comparing the average message delay, for a given set of traffic co nditions, for a hypercube network and a mesh network. Hypercube networ ks may involve more links than meshes. How do you incorporate that hyp ercube networks may require more complex hardware? Should the total pa th width of all the links the networks employ be the same, Or should t he two networks require the same number of transistors per switch? Thi s article explores the problems of determining which metrics or weight ed set of metrics designers should use to compare networks and how the y should apply these metrics to yield meaningful information. The auth ors also look at problems in conducting fair and scientific evaluation s.