Two-Level Adaptive Branch Prediction

ECE7660 Reading Assignment 2

Analysis and Critique of Reading Assignment 2 Paper “Alternative Implementations of Two-Level Adaptive Branch Prediction”

-Pavan Bhujannavar

In this paper we authors have proposed a new dynamic branch predictor (Two-Level Adaptive Branch Prediction) by which substantially higher accuracy is obtained when compared to any other schemes.Using a high level language, application is written as a program and executed sequentially. Sequential model execution of the program makes use of instruction sets, execution of Instructions one by one and if not done properly leads to Interrupts, exceptions and Branch misprediction. In order to support better performance processors by minimizing the overhead associated with mispredicted branches during Parallel Execution i.e. to execute more than one instruction simultaneously and control, Two-level Adaptive Branch Prediction was proposed, which alters the branch prediction algorithm. Thus achieving greater degrees of instruction level parallelism. Branches change the normal flow in unexpected ways which interrupts the normal fetch and issue operation of instructions as to resolve the outcome of branches considerable numbers of cycles are needed. Branch prediction technique is employed with an effort to determine which particular program branch will be taken for execution. Appropriate branch prediction lets the processor predict a program's flow of control and, therefore, only fetch those instructions on the program's actual path of execution. If mispredicted, the work must be redone and the correct instruction needs to be fetched. Static branch prediction relies on the information associated prior to runtime whereas in the dynamic branch prediction, processor tries to e...

... middle of paper ...

...t, context switch and effects of various history register lengths- due to the increasing interference in the branch history the IHRT scheme performs the best and according to the simulations, it is evident that increasing the history register length often improves the prediction accuracy and effect of context switch- uses the BHT to keep track of branch history, the table needs to be flushed during a context switch are portrayed. On comparison, best predictors gave 94.4% accuracy whereas two-level predictors pushed it to 97%, the other 3% was still substantial miss rate and current version of simple scalar gives 98.8% hit rate.

Two-level adaptive branch predictors provide higher level of branch prediction accuracy and have been a main focus area of branch prediction research since 1991 the year in which Yeh and Patt proposed the Two-Level Adaptive Branch Predictor.