CAREER: Slice-Processing: Highly-Accurate Prediction for Future High-Performance Processors

Northwestern University, Evanston IL

Investigators

Abstract

DATE: April 19, 2000 PROPOSAL NUMBER: C-CR 9984371 INSTITUTION: Northwestern University PI: Andreas Moshovos TITLE: CAREER: Slice-Processing: Highly-Accurate Prediction for Future High-Performance Processors ABSTRACT Prediction-driven speculative techniques are used in virtually all modern, high- performance processors. Such techniques offer a promising direction for further boosting performance. However, the benefits so obtained are only as good as the underlying prediction-based methods are at guessing what the program needs. This project investigates slice-prediction, a paradigm that aims at providing highly accurate prediction beyond what is possible with the existing methods. Existing methods are primarily outcome-based since they rely on repeating patterns in a program's outcome stream (e.g., addresses or branch directions). However, such patterns do not always exist. More importantly, as current outcome-based methods are perfected, irregular patterns increasingly dominate performance. Slice-prediction attacks such cases by predicting the computation- slice that produces such irregular yet performance-critical outcomes. Predicted slices are then treated as miniature, autonomous programs, which can precompute performance critical outcomes. The intuition is that while the outcome of a program can be irregular the method used is typically relatively simple and fairly stable. This project focuses on architecturally invisible, hardware implementations of slice-prediction, and targets the following applications: (1) Branch Prediction and (2) Data Prefetching. Performance achieved by slice-prediction is evaluated through a detailed software simulator over a set of widely used benchmarks.