IO latency hiding in pipelined architectures

This paper reports upon development of a novel mathematical formalism for analyzing data pipelines. The method accounts for IO and CPU latencies in the stages of the data pipeline. An experimental pipeline was constructed using a video encoder, frame processing, and transport of the frames over an IP (Internet protocol) network. The pipelined architecture provides a method to overlap processing with DMA, encoding and network transport latency so that streams can be processed with optimal scalability. The model expectations were compared with experimental test results and found to be consistent. The model is therefore expected to provide a good estimate for the scalability of streaming video-on-demand systems. Video-on-demand is a rapidly growing service segment for entertainment, advertising, on-line education, and a myriad of emergent applications.