The development of a CHAID-based model for CHITRA93

Abstract

The complexity of the behavior of parallel and distributed programs is the major reason for the difficulties in the analysis and diagnosis of their performance. Complex systems such as these have frequently been studied using models as abstractions of such systems. By capturing only the details of the system which are considered essential, a model is a replica of the complex system which is simpler and easier to understand than the real system. CHITRA92, the second generation of the performance analysis tool CHITRA, builds a continuous time semi-Markov chain to model program behavior. However, this model is limited to representing relationships between states which are only immediate predecessors or successors of each other. This project introduces and implements a new empirical model of the behavior of software programs which is able to represent dependencies between nonsequential program states. The implementation combines deterministic and probabilistic modeling and is based on the Chi Automatic Interaction Detection (CHAID) statistical technique designed for investigating categorical data. The empirical model, constructed by analyzing an ensemble of program execution sequences, is stochastic and non-Markovian in the form of an N -step Transition Matrix. The algorithm is integrated as one of the modeling subsystems of CHITRA93, the third generation of CHITRA.