Kaizen Programming for predicting numerical linear algebra operations performance

Abstract

Perhaps the most critical sparse linear algebra operation is the product of a sparse matrix and a dense vector, known as SpMV. This motivates the high-performance computing community to make a continuous effort to produce efficient implementations of this kernel for the most widespread parallel computing platforms. There are numerous implementations for the SpMV, spanning different algorithms and sparse matrix representations, with a wide spectrum of GPU SpMV routines. It is interesting to provide means to automatically select the implementation that will likely provide the best performance for a given matrix. In the present work we evaluate the use of Kaizen Programming (KP) to build classifier models. We worked with two classifiers based on KP in order to select the best routine based on eight sparse matrix characteristics. We found that both approaches build good classifiers, with almost 74 and 83% of accuracy, respectively to each approach.