Using Structured Modelling for Efficient Performance Prediction of Parallel Systems.
Markus Siegle · 1993
: A method for analyzing parallel systems with the help of structured Markovian models is presented. The overall model is built from interdependent submodels, and symmetries are exploited automatically by grouping similar submodels in classes. During model analysis, this leads to a state space reduction, based on the concept of exact lumpability. 1. Introduction The high complexity of today's parallel computers --- due to the large number of concurrently active and mutually dependent hardware and software components --- makes them difficult to understand for human users. Developing efficient software for this class of machines is therefore very time-consuming and expensive. The resulting performance often fails to meet expectations, therefore requiring major re-implementation. Event-based performance models can be used to alleviate this problem. They help humans to understand the dynamic behaviour of parallel systems by abstracting the overwhelming number of details of the real world...