Early Design Phase Cross-Platform Throughput Prediction for Industrial Stream-Processing Applications

Industrial embedded platforms are often used to execute stream-processing applications, from which the results are used by actuators. On average, these stream-processing applications should at least meet the required throughput of their actuators, which poses a real-time requirement on the system. To avoid extra costs and delays, it is desired to estimate during the early design phase if a combination of an embedded platform and a stream-processing application can achieve the required throughput. The throughput of a stream-processing application executed on different embedded platforms can be predicted by modeling them using static or measurement based analysis. However, during the early design phase it can be desirable to have a model that allows a large set of embedded platforms to be considered, where embedded platforms with predictive instructions are supported. This paper presents a gray-box approach applicable during the early design phase to perform cross-platform throughput predictions for industrial stream-processing applications and their embedded platforms. A three step regression-based approach is presented, which uses an expression based on Amdahl’s law for the discrete scaling of workload over cores and a large database with CPU performance scores to perform cross-platform throughput predictions. Validation, with a limited set of platforms, showed the usability of the approach. The pragmatic approach is based on a prototype industrial digital image processing application for a printer from Oce, which is also used to present the approach.