An Algorithm for O { Line Detection of Phasesin Execution Pro les ?

This paper describes a method for the detection of phases in time{series data. The speciic application considered is time{series data which describes an execution proole of a parallel program. The algorithm is described and examples are presented. This paper presents the problem in the broader sense of general oo{line time{series analysis. In analyzing time{series data, oo{line algorithms enjoy the luxury of more compu-tationally intensive approaches than possible with on{line algorithms. This paper describes an oo{line method for the analysis of time{series data which detects periods of maximum homogeneity in time{series data, speciically phases in execution prooles. The technique is applied to execution prooles from program executions on hypercube computers. This is used to provide another means of workload characterization, and has been used to predict the speedup of parallel programs.