QoS-aware Resource-utilisation Self-adaptive (QRS) Framework for Distributed Data Stream Management Systems
Date
Authors
Advisors
Journal Title
Journal ISSN
ISSN
DOI
Volume Title
Publisher
Type
Peer reviewed
Abstract
The last decade witnessed a vast number of Big Data applications in the science and industry fields alike. Such applications generate large amounts of streaming data and real-time event-based information. Such data needs to be analysed under the specific quality of service constraints, which must be done within extremely low latencies.
Many distributed data stream processing approaches are based on the best-effort QoS principle that lack the capability of dynamic adaptation to the fluctuations in data input rates. Most of the proposed solutions tend to either drop some of the input data (load shedding) or degrade the level of QoS provided by the system. Another approach is to limit the data ingestion input rate using techniques like backpressure heartbeats, which can affect the worker nodes that causes an output delay. Such approaches are not suitable to handle certain types of mission-critical applications such as critical infrastructure surveillance, monitoring and signalling, vital health care monitoring, and military command and control streaming applications.
This research presents a novel QoS-aware, Resource-utilisation Self-adaptive (QRS) Framework for managing data stream processing systems. The framework proposes a comprehensive usage model that encompasses proactive operations followed by simultaneous prompt actions. The simultaneous prompt actions instantly collect and analyse the performance and QoS metrics along with running data streams, ensuring that data does not lose its current values, whereas the proactive operations construct the prediction model that anticipate QoS violations and performance degradation in the system. The model triggers essential decision process for dynamic tuning of resources or adapting a new scheduling strategy. A proof of concept model was built that accurately represents the working conditions of the distributed data stream management ecosystem. The proposed framework is validated and verified. The framework’s several components were fully implemented over the emerging and prevalent distributed data streaming processing system, Apache Storm.
The framework performs accurate prediction up to 81% about the system’s capacity to handle data load and input rate. The accuracy reaches up to 100% by incorporating abnormal detection techniques. Moreover, the framework performs well compared with the default round-robin and resource-aware schedulers within Storm. It provides a better ability to handle high data rates by re-balancing the topology and re-scheduling resources based on the prediction models well ahead of any congestion or QoS degradation.