University of Houston
Department of Computer Science

In Partial Fulfillment of the Requirements for the Degree of
Master of Science

Reshma Khatri

Will defend her thesis

Using Artificial Neural Networks to Select the Best File Access Prediction Heuristic

Abstract

Conventional file access predictors use a single heuristic such as Last Successor or Recent Popularity to predict which file a computer system will access next. We introduce a dynamic predictor that runs several conventional file access predictors—Last Successor, Stable Successor, and Recent Popularity—in parallel. It employs a neural network to learn over time which predictor is currently the most accurate for a given pattern of file accesses and then selects that predictor to make the next prediction. We were able to significantly improve upon the accuracy, success-per-reference, and effective-success-rate-per-reference of Recent Popularity by using our neural network-based file access predictor with the proper tuning. In particular, we generated incorrect predictions 43.63% to 53.11% less often than Recent Popularity when using our predictor with a standard configuration. With manual tuning for each trace, we were able to improve upon the misprediction rate of Recent Popularity by 69.02% to 78.43%. Due to our high accuracy, we achieved an effective-success-rate-per-reference 5.72% to 14.71% higher than that of Recent Popularity using our standard configuration, which proved to be equivalent to that found by manual tuning for each trace.

Date: Friday, November 3, 2006
Time: 3:00 PM
Place: 550 PGH

Faculty, students, and the general public are invited.
Advisor: Prof Jehan-François Pâris