Process Switches and Branch Prediction Accuracy

There were several hypotheses that motivated this research project. First, we proposed that running multiple processes pollutes the history in branch predictors due to aliasing. This pollution decreases accuracy. Second, we argued that branch prediction accuracy can be improved over existing techniques by removing this aliasing. Specifically, the history table can be partitioned to allow each process, or a subset of the most frequently executed processes, to have their own history. This project successfully shows that aliasing in the branch predictor history does occur due to multiple processes executing, and quantifies the frequency of that problem. The accuracy of existing predictors can be improved, and adding separate history tables is an effective method of removing the aliasing problem. Specifically, a prediction improvement of 0.5 – 3% can be obtained through this technique, and this improvement is always in addition to any prediction algorithm that suffers from aliasing. The actual accuracy penalty due to aliasing is much higher than this improvement would seem to suggest. The relative infrequency of process switches, however, and the short time of approximately 2000 branches to “repair” the destructive aliasing in the branch history table after a process switch, both combine to significantly limit the potential for improvement via this technique. When the hardware overhead of multiple history tables is taken into account, this technique is not recommended for generalpurpose applications