Performance Curves for better Performance Predictions of Parallel Applications in Multicore Environments

Model-based performance prediction for parallel applications on architectural models suffers from significant inaccuracies. 
A major reason is that current model-based performance prediction approaches consider CPU speed as a single metric for multicore performance. 

Thus, in this paper, we investigate performance-influencing factors for multicore environments, execute extensive experiments to determine their impact on the performance, and extract performance curves for characteristic behaviours.

As a result, we present a set of performance curves to software engineers which enables them to increase the performance prediction power in an easy to use manner. 
Further, we evaluate the approach using 13 SPEC Benchmarks and could show that our approach reduces the prediction error by up to 60% and increases the accuracy by up to 98% for certain scenarios.