Workloads, observables, benchmarks and instrumentation
Gordon Lyon, Robert Snelick · 1990
Main emphasis is upon a compact user-level summary that captures the performance variabilities of a system.A dependency tree provides a clear, static declaration of the relationships among a very limited number of major system resources that explain most performance variance.This approach is especially effective with hardware instrumentation that decouples workload from observation, for otherwise a compact set of observables may be unavailable.The tree supports simple predictions and promotes more meaningful comparisons of workloads.Accounting for the sources of performance variation shown in the tree can inspire new methods of assessment; a "time dilation" technique illustrates this for loosely- coupled systems with local clocks.