Dynamic Resource Management of Network-on-Chip Platforms for Multi-stream Video Processing

This thesis considers resource management in the context of parallel multiple video stream decoding, on multicore/many-core platforms. Such platforms have tens or hundreds of on-chip processing elements which are connected via a Network-on-Chip (NoC). Inefficient task allocation configurations can negatively affect the communication cost and resource contention in the platform, leading to predictability and performance issues. Efficient resource management for large-scale complex workloads is considered a challenging research problem; especially when applications such as video streaming and decoding have dynamic and unpredictable workload characteristics. For these type of applications, runtime heuristic-based task mapping techniques are required. As the application and platform size increase, decentralised resource management techniques are more desirable to overcome the reliability and performance bottlenecks in centralised management. In this work, several heuristic-based runtime resource management techniques, targeting real-time video decoding workloads are proposed. Firstly, two admission control approaches are proposed; one fully deterministic and highly predictable; the other is heuristic-based, which balances predictability and performance. Secondly, a pair of runtime task mapping schemes are presented, which make use of limited known application properties, communication cost and blocking-aware heuristics. Combined with the proposed deterministic admission controller, these techniques can provide strict timing guarantees for hard real-time streams whilst improving resource usage. The third contribution in this thesis is a distributed, bio-inspired, low-overhead, task re-allocation technique, which is used to further improve the timeliness and workload distribution of admitted soft real-time streams. Finally, this thesis explores parallelisation and resource management issues, surrounding soft real-time video streams that have been encoded using complex encoding tools and modern codecs such as High Efficiency Video Coding (HEVC). Properties of real streams and decoding trace data are analysed, to statistically model and generate synthetic HEVC video decoding workloads. These workloads are shown to have complex and varying task dependency structures and resource requirements. To address these challenges, two novel runtime task clustering and mapping techniques for Tile-parallel HEVC decoding are proposed. These strategies consider the workload communication to computation ratio and stream-specific characteristics to balance predictability improvement and communication energy reduction. Lastly, several task to memory controller port assignment schemes are explored to alleviate performance bottlenecks, resulting from memory traffic contention.