Performance Prediction through Time Measurements
Roman Iakymchuk · 2011
Abstract. In this article we address the problem of predicting performance of linear algebra algorithms for small matrices. This approach is based on reducing the performance prediction to modeling the execution time of algorithms. The execution time of higher level algorithms like the LU factorization is predicted through modeling the computational time of the kernel linear algebra operations such as the BLAS subroutines. As the time measurements confirmed, the execution time of the BLAS subroutines has a piecewise-polynomial behavior. Therefore, the subroutines time is modeled by conducting only few samples and then applying polynomial interpolation. The validation of the approach is established by comparing the predicted execution time of the unblocked LU factorization, which is built on top of two BLAS subroutines, with the separately measured one. The applicability of the approach is illustrated through performance experiments on Intel and AMD processors.