Performance modeling of parallel systems with flexible precision

The spectrum of approaches to performance prediction of parallel systems ranges from low-cost, low-precision compile-time methods to high-cost, high-precision methods with an equal variety of underlying modeling formalisms (e.g., task graphs, queueing networks, simulation languages) which are mutually incompatible. A performance prediction technique is described which unifies a number of existing approaches within one formalism. As the analysis is based on a model reduction calculus the approach allows for a flexible trade-off between precision and cost, effectively ranging from simulation to static analysis