Performance Prediction of Parallel CPU and GPU Applications Using Fractals
Rodrigo Escobar, Rajendra V. Boppana · 2018
Accurate estimates of parallel application runtimes can be used for better job scheduling and utilization of high performance computing resources, application code finetuning, power consumption, and improved response time. We present a performance model using the theory of fractals to estimate the runtimes of parallel CPU and GPU applications. The fractal model is suitable for applications whose major computational phases' runtimes can approximated by polynomials of the form aNc, for input size N and some constants a and c. Our approach requires only a few small-scale runs of the target application and the analysis of the corresponding runtimes to predict the application runtime for a larger input. Our method avoids more laborious approaches, such as static analysis of source code or capturing and analyzing detailed execution trace logs, proposed in the literature. Our experimental results for 21 well-known parallel scientific applications on CPUs and GPUs show that this method has good prediction accuracy with errors of less than 12% in most cases.