Metrics for Metrics: Why It Is Difficult to Compare Interconnection Networks Or How Would You Compare an Alligator to an Armadillo?
Howard Jay Siegel, John K. Antonio, Kathy J. Liszka · 2005
Difficulties associated with metric-based design and evaluation of interconnection networks, in the context of massively parallel processing systems, are discussed from a relatively philosophical perspective. The issuesTywhich metric or weighted set of metrics to use@ehow to meaningfully apply metrics“46“are central themes throughout the paper. It is difficult to isolate meaningful comparative network performance measurements when using application benchmarks on different systems. Furthermore, the strong interdependence among application, system, and network parameters makes it difSlcult to apply the classical experimental method, which is to “measure” the effect of changing one parameter while holding all other parameters constant. The issue of fairness is discussed in the context of the classical experimental method, and one aspect of fairness related to algorithm mapping is illustrated through an example.