Predicting the Performance of Automatically Scalable Computation (ASC)

Abstract

The Automatically Scalable Computation (ASC) architecture is a new approach to automatic parallelization that transforms parallelization into a machine learning problem. The underlying principle is that if we observe the state of the machine repeatedly at a given place in a sequential program, we can build a model to predict the state of the machine at that place over time. The ``place" in the program is defined by an address loaded into the instruction pointer (IP). We present a machine learning approach to automatic IP selection. Our approach relies on the observation that the error function of ASC's internal machine learning model decreases over time for a good IP value. For a given program and IP value, we build a predictive model for ASC's error function value over time. The inputs are a program and IP value and the target is ASC's error function value. We present methods for representing the program and IP value as a feature set. We also present and evaluate three different machine learning models for predicting the error function values.