Towards low-cost, high-accuracy classifiers for linear solver selection

The time to solve linear systems depends to a large extent on the choice of the solution method and the properties of the coefficient matrix. Although there are several linear solution methods, in most cases it is impossible to predict apriori which linear solver would be best suited for a given linear system. Recent investigations on selecting linear solvers for a given system have explored the use of classification techniques based on the linear system parameters for solver selection. In this paper, we present a method to develop low-cost high-accuracy classifiers. We show that the cost for constructing a classifier can be significantly reduced by focusing on the computational complexity of each feature. In particular, we filter out low information linear system parameters and then order the remaining parameters to decrease the total computation cost for classification at a prescribed accuracy. Our results indicate that the speedup factor of the time to compute the feature set using our ordering can be as high as 262. The accuracy and computation time of the feature set generated using our method is comparable to a near-optimal one, thus demonstrating the effectiveness of our technique.