Behavior-Based branch prediction by dynamically clustering branch instructions

Hans Vandierendonck, V. Desmet, K. De Bosschere

Research output: Contribution to journalArticlepeer-review

1 Citation (Scopus)

Abstract

Conditional branches frequently exhibit similar behavior (bias, time-varying behavior,...), a property that can be used to improve branch prediction accuracy. Branch clustering constructs groups or clusters of branches with similar behavior and applies different branch prediction techniques to each branch cluster. We revisit the topic of branch clustering with the aim of generalizing branch clustering. We investigate several methods to measure cluster information, with the most effective the storage of information in the branch target buffer. Also, we investigate alternative methods of using the branch cluster identification in the branch predictor. By these improvements we arrive at a branch clustering technique that obtains higher accuracy than previous approaches presented in the literature for the gshare predictor. Furthermore, we evaluate our branch clustering technique in a wide range of predictors to show the general applicability of the method. Branch clustering improves the accuracy of the local history (PAg) predictor, the path-based perceptron and the PPM-like predictor, one of the 2004 CBP finalists.
Original languageEnglish
Pages (from-to)919-931
Number of pages13
JournalJOURNAL OF INFORMATION SCIENCE AND ENGINEERING
Volume24
Issue number3
Publication statusPublished - May 2008

ASJC Scopus subject areas

  • Information Systems

Fingerprint

Dive into the research topics of 'Behavior-Based branch prediction by dynamically clustering branch instructions'. Together they form a unique fingerprint.

Cite this