Increasing Efficiency of Data-flow Based Middleware Systems by Adapting Data Generation

Many data-flow oriented applications are based on the pipe-and-filter concept. This paper presents an improvement of the state of the art for multi-threaded pipe-and-filter processing architectures. We present a novel approach for adapting the time of data generation in the pipeline where adjacent pipeline stages exchange information about the current utilization. We compare our approach to a traditional input data driven pipeline and achieve a significant reduction of the processing delay and required memory consumption. The improvement of the presented system is based on self-adapting the data generation rate in the processing pipeline. This adaptation results in two key efficiency improvements: (i) the reduction of the time data elements spend in the pipeline and (ii) the reduction of the memory requirement for communication buffers. These improvements are of special interest for reactive and interactive multi-camera applications where short delays of the image processing pipelines is often required. The presented approach enables any data-flow based application to execute with reduced memory usage, reduced execution delay and the highest possible data-rate.