Performance Prediction/Verification Scheme for Data-driven Realtime Systems
Yasuhiro Wabiko, Hiroaki Nishikawa · Parallel and Distributed Processing Techniques and Applications · 2002
This paper presents the effectiveness of the performance prediction/verification scheme to interactively support development of realtime systems using super-integrated dynamic data-driven processors CUE (Coordinating Users’ requirements and Engineering constraints). Since a realtime system must satisfy performance requirements such as turnaround time and data flow rate, performance prediction/verification is crucial to guarantee that a realtime system meets such requirements. The authors have proposed Realtime Execution System for CUE-series data-driven processors (RESCUE). CUE processors have ideal multi-processing capability without any interference among independent processes running simultaneously. Hence, the runtime performance of CUE processor can be predicted by static analysis of the program, excluding dynamic behavior of elastic pipeline comprising a CUE processor. This dynamic behavior can be verified by pipeline-level emulator for CUE processors. The authors have implemented performance prediction/verification scheme on RESCUE. This paper first discusses the accuracy of the prediction results. Through this discussion, it is shown that our prediction scheme can correctly predict runtime performance of CUE processor system, except dynamic behavior caused by wait-matching of two operands. Such a dynamic behavior can be completely verified by pipeline-level emulator for CUE processors. This paper then discusses usability of the performance prediction/verification facility of RESCUE, by applying RESCUE to the development of protocol handler, audio compression and media processing. As a result, RESCUE can support not only detection of the bottleneck of a realtime system based on CUE processors, but also tuning of CUE processor architecture.