An Integrated Performance Estimation Approach in a Hybrid Simulation Framework

The increasing complexities of today’s embedded multimedia and wireless devices have ushered in the era of heterogeneous MultiProcessor System-on-Chip (MPSoC) architectures. This trend, in turn, have made software parallelization and optimization a subject of utmost importance for today’s systems. Nowadays, providing efficient software implementations is not only mandatory for the final products, but also necessary for Design Space Exploration (DSE) of the numerous hardware choices available for an MPSoC development. Unfortunately, such co-exploration of different software solutions and hardware architectures usually requires an extraordinarily large effort due to the continually increasing gap between the speed of the real processor hardware and that of the available instruction set simulators. This problem can be greatly alleviated by using cycle approximate, but fast simulation models for early DSE where relative merits of different design solutions are more important than hundred percent cycle accuracy. To address the issue of fast performance estimation for DSE, HySim a hybrid simulation framework which consists of an Instruction Set Simulator (ISS) and a native execution engine called Virtual CoProcessor (VCP) has been proposed. Virtualization is performed so that parts of the execution can be shifted to the native engine without sacrificing the functional correctness of the whole application. High execution speed with performance estimation is available at the VCP side, and therefore, combined with the ISS, good accuracy can be obtained. Our previous work introduced two performance estimation approaches annotation based and dynamic profiling based. However, some new questions are opened on how to combine these two approaches, and how to effectively partition an application to VCP and ISS. In this paper, annotation based performance estimation is used to facilitate dynamic profiling, and a preliminary sampling approach is also introduced to open the possibility of further reducing dynamic profiling overhead by inter/extrapolation.