TR-111: Impact of Prediction Accuracy on the Performance of a Pipeline Computer

Abstract

A general weakness of Pipeline Computers is ‘bubbles’ in the pipeline due to conditional branching. Most of the solutions to the branch problem attempt to predict whether or not a branch will be taken. If the prediction is correct, then initiation of the correct sequence of instructions can continue without delay. In this report, the performance of a Pipeline Processor is analysed to give a measure of the efficiency or utilisation of pipeline segments in terms of the probability of correct branch prediction (referred to as the prediction accuracy).
The analysis is based on the space-time relationship. The performance of a Pipeline Processor with unequal segment times is analysed noting that the two types of conditional branch instructions in the instruction sets of computers have different effects on the performance. The result of the analysis enable one to study the sensitivity of the performance of a Pipeline Processor to the number of conditional branch instructions and the probability of correct branch prediction as well as to estimate the utility of the different branch prediction strategies that have been proposed recently [2]. Although it is extremely difficult to obtain the value of 1 for prediction accuracy, this study reveals that a value of 0.8 can give reasonably good values of utilisation.