Branch Meta-Prediction by Experts

Predicting the direction of conditional branches with high accuracy has become critical to the success of new deeper and wider processor pipelines. However, choosing a branch predictor is hard because the best choice is workload-dependent. Meta(or Hybrid-) predictors provide a rudimentary FSA-based heuristic for choosing between two predictors on-line. Such approaches have little or no theoretical motivation, and offer no performance guarantees. Recent developments in the Expert Framework provide explicit mechanisms for combining a pool of predictors and come with powerful theoretical results, bounding the loss the of master (meta-) predictor in terms of the loss of predictors in its pool. This paper shows how an arbitrary number branch predictors can be used as experts, and how the Expert Framework can be used to construct a multipredictor. Our expert-based meta-predictors can achieve up to 62% fewer misses than the best predictor chosen in hindsight.