A Statistical Method for Real-Time
Software Performance Estimation

Patrick Facchini

Dept. of Computer and Information Science, Linköping University, Sweden


This report is the result of an equivalent to a Masters thesis project done at the Linköping University, Sweden, during the period February to July 1996. The execution time of a program or a block of a program is important in real-time design. Different methods to estimate the performance of a program, taking into account the features of a processor (clock frequency, memory access time), have been formulated. However, those methods cannot be used early in the design. This report introduces a new method of estimation based on statistical analysis and measurements which model the hardware. It uses linear regression to estimate the execution time of a program on different systems. A model, based on the execution time of narrow programs (describing basic operations), determines the characteristics of the different processors. Then, the estimation of the execution time of a program is the product of the values of this model and the regression parameters, which are calculated with linear regression. The method has been tested with Ada programs on Sparcstations and PCs. We point out the importance of the accuracy of measurements, which may have a non-negligible influence on the results. The error of the estimates is very small for a set of only Sparcstations, but is larger when PCs are added to that set. We discuss plausible causes of these results. The first part of this report is a description of the subject and of previous research. Then the report deals with the method of evaluation, followed by the description of the experiments. The last part is the analysis and conclusion about this method.

Download the paper (102483 bytes).