| Abstract: |
The exponential growth of Internet of Things (IoT) devices has led to a massive volume of time-sensitive data that need real-time processing at the network edge. Cloud-centric incumbent architectures are not only imposing unacceptable time delays for applications such as autonomous vehicles, industrial automation and augmented reality, but also saturating network bandwidth to relay infinite amount of data. Edge computing is an alternative paradigm that processes data closer to the source, and existing frameworks have limitations in dynamic resource assignment, heterogeneous device administration and work distribution within a distributed edge infrastructure. In this paper we present DeepEdge, an intelligent distributed computing framework that tackles these challenges through the following key innovations: (i) A deep reinforcement learning-based resource orchestrator achieving 43.7% reduction in task completion time by anticipative workload placement and continual resource scaling; (ii) a hierarchical computation offloading algorithm that minimizes deadline-violation rates considering both latency bounds, energy consumption constraints and computational complexity for tiered device-edge-cloud distribution of tasks and (iii) an integrated predictive maintenance solution employing temporal convolutional networks to predict device failures with 94.2% accuracy allowing proactive reallocation of resources. Comprehensive experiments on a 500 node testbed with scenarios including smart city, industrial IoT and healthcare monitoring show that DeepEdge reduces end-to-end latency by 67.3% when comparing against cloud-only approaches and by 34.8% with respect to state-of-the-art edge frameworks whilst achieving an energy efficiency gain of 41.2%. The system can handle more than 2.3M events per second with less than 10ms latency at the p99 horizon, which sets new standards for the distributed IoT analytics. |
