OPTS: increasing branch prediction accuracy under context switch

Accurate branch prediction is essential for obtaining high performance in pipelined superscalar processors. Though many dynamic branch predictors have been proposed to obtain high prediction accuracy, they cannot perform as expected under context switches. It is observed that context switches, even at fairly large intervals, can seriously degrade the performance of dynamic branch predictors. In this paper we measure the effect of context switch on branch prediction, and present a new scheme which saves and restores branch predictor table when a context switch occurs. This scheme takes advantage of multiple small predictors which preserve branch predictor tables of independent processes. The effectiveness of reducing interprocess interference is evaluated by simulations.