Machine learning for microarchitectural prediction
Document typeConference report
PublisherBarcelona Supercomputing Center
Rights accessOpen Access
Cache replacement and branch prediction are two important microarchitectural prediction techniques for improving performance. We propose a data-driven approach to designing microarchitectural predictors. Through simulation, we collect traces giving detailed control-flow and memory behavior. Then use stochastic search techniques, such as genetic algorithms, to find points in a large design space of predictors that yield good accuracy on the traces. We then evaluate the predictors on held-out data. This talk will present two techniques resulting from this methodology. In Multiperspective Branch Prediction, many features and their parameters are tuned using a genetic algorithm to yield a very accurate perceptron-based branch predictor. Multiperspective Reuse Prediction uses the same idea for cache management. Many features of memory accesses to predict the reuse of a given memory access. The features and their parameters are chosen by a stochastic search yielding a very accurate predictor. This predictor is applied to a placement, replacement, and bypass optimization that out-performs the state of the art.
CitationJiménez, D. A. Machine learning for microarchitectural prediction. A: 3rd Severo Ochoa Research Seminar Lectures at BSC, Barcelona, 2016-2017. "Book of abstracts". Barcelona: Barcelona Supercomputing Center, 2017, p. 49-50.