Advanced Computer Architecture, Imperial College 2001
19
Warm-up effects and context-switching
•In real life, applications are interrupted and some other program runs for a while (if only the OS)
•This means the branch prediction is regularly trashed
•Simple predictors re-learn fast
–in 2-bit bimodal predictor, all executions of given branch update same 2 bits
•Sophisticated predictors re-learn more slowly
–for example, in (2,2) gselect predictor, prediction updates are spread across 4 BHTs
•Selective predictor may choose fast learner predictor until better predictor warms up
•