Predicting Algorithm Performance

We examine the problem of predicting the time that an algorithmwill take to solve a problem of speci ed size on a speci ed computer, and of predicting the uncertainty in the estimate. The predictions are based on an adaptive sampling approach, constrained by a xed time budget for all experiments. Complications arise due to the poorly understood nature of system variation, because small inputs may be poor predictors of the performance for large inputs due to changes in cache behavior, and because time is both a constraint and the variable being estimated.