Simulation-driven Software Performance Estimation for Fast Design Space Exploration

In this paper, we propose a novel performance estimation method of software function blocks considering the effect of architecture variation, compiler optimization, and data dependent behavior. In the proposed design space exploration framework, a system behavior is specified as a composition of function blocks and the execution order of the blocks is known a priori. We run the entire application with code augmentation on the instruction set simulator of the target processor to obtain the accurate performance information of individual function blocks. In the design space exploration loop, the performance of the entire application is easily computed as a linear combination of function block performance values. Experimentation with real-life applications proves the viability and efficiency of the proposed technique.