Architecture for Predicting Live Video Transcoding Performance on Docker Containers

Video can be streamed live with different applications (e.g. YouTube Live, Periscope). Typically, the video content is adapted for end users based on receiving client's capabilities, and network bandwidth. The adaptation is realized with different video representations, which are created by transcoding the original video content. When video is streamed live, transcoding has to be completed within real time constraints, which is a computationally demanding process. Particularly, live transcoding should be enabled efficiently by a content distributor to minimize resource provisioning costs. The contribution of this paper is an architecture for predicting live video transcoding performance on a Docker-based platform. Particularly, cloud resource management for live video transcoding has been focused on. A model was trained based on measurements in different transcoding configurations. Offline evaluation results indicate that live transcoding speed or CPU usage can be predicted with 3-8 % accuracy. When video is transcoded on virtual machines based on predictions in a prototype system (live), live transcoding speed prediction accuracy is within a similar range as the offline performance, but worse for CPU usage prediction (5-15%). In most cases the specified range for transcoding speed and CPU usage can be achieved at least with a precision of 76%.