Michele Co

Potential for Branch Predictor Adaptation at the Program and Phase Level for Performance and Energy-Efficiency

M. Co, D. A.B. Weikle, and K. Skadron.
University of Virginia, Department of Computer Science Technical Report No. CS-2005-19

Abstract
Experiments to determine the potential for program-level and/or phase-level adaptation of the branch predictor configuration for the purpose of total processor energy savings were performed. The performance and energy-efficiency of an 8-wide issue, out-of-order processor with six different branch predictors were evaluated on the SPECcpu2000 benchmark suite. Each branch predictor was compared to the branch predictor with the highest overall average IPC. The comparison was performed at both the program-level and at 1M instruction intervals to determine the potential of adaptation of the branch predictor configuration to improve overall processor energy-efficiency. The results of these experiments indicate little potential for branch predictor adaptation for the SPECcpu2000 benchmark suite using the branch predictors evaluated. Additional results show that the piecewise linear branch predictor consumes significantly more energy than its close-performing competitors and that the hashed perceptron predictor is comparable both in performance and energy-efficiency to the overall best-performing branch predictor in this study, O-GEHL.


Available in postscript or pdf