Optimal Hardware Parameters Prediction for Best Energy-to-Solution of Sparse Matrix Operations Using Machine Learning Techniques

• https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=4&ved=2ahUKEwj45dbYkMfdAhUilYsKHQe1DkwQFjADegQIBxAC&url=https%3A%2F%2Fwww.thinkmind.org%2Fdownload_full.php%3Finstance%3DINFOCOMP%2B2018&usg=AOvVaw0F5eFy3SoDGqt3wTWnO1GV

sparse, neural networks, energy efficiency, prediction

Combinations of 3 hardware parameters (number of threads, core and uncore frequency) were tested for each of the 4 sparse algorithms (matrix-matrix addition, matrix-matrix multiplication, matrix-vector multiplication in IJV and CSR format) on a set of several thousands matrices for the purpose of identifying the best energy-to-solution setting for each matrix and sparse operation. On this set of data, the possibility of optimal hardware setting prediction based on the properties of each matrix were analysed for each sparse algorithm. A calculation of Pearson correlation coefficient between the matrices' properties and optimal hardware parameters showed no direct correlation (highest 0.33 for x-y, lowest -0.25 for a-b). A neural network with back-propagation learning was used for deeper analysis to see if matrix properties correspond to hardware settings. The input neurons represented properties of given matrix, output neurons represented optimal hardware parameters. Network properties (hidden neurons per layer, hidden neuron layers, learning coefficient and learning strategy) impact on prediction accuracy were analysed and the results showed