Predicting Program Behavior Based On Objective Function Minimization

Abstract

Computer systems increasingly rely on dynamic management of their operations with the goal of optimizing an individual or joint metric involving performance, power, temperature, reliability and so on. Such an adaptive system requires an accurate, reliable, and practically viable metric predictors to invoke the dynamic management actions in a timely and efficient manner. Unlike ad-hoc predictors proposed in the past, we propose a unified prediction method in which the optimal metric prediction problem is considered as that of minimizing an objective function. Choice of the objective function and the model type determines the form of the solution whether it is a closed form or one that is numerically determined through optimization. We formulate two particular realizations of the unified prediction method by using the total squared error and accumulated squared error as the objective functions in conjunction with autoregressive models. Under this scenario, the unified prediction method becomes linear prediction and the predictive least square (PLS) prediction, respectively. For both of these predictors, there is a analytical closed form solution that determines model parameters. Experimental results with prediction of instruction per cycle (IPC) and L1 cache miss rate metrics demonstrate superior performance for the proposed predictors over the last value predictor on SPECCPU 2000 benchmarks where in some cases the mean absolute prediction error is reduced by as much as 10-fold.