Bandit-based Delay-Aware Service Function Chain Orchestration at the Edge [article]

Lei Wang, University Of Calgary, Majid Ghaderi
2021
Mobile Edge Computing (MEC) enables both cloud computing and edge computing for mobile users, providing them with intensive computing resources and proximity to the data sources. When combined with network function virtualization (NFV), MEC provides users with promising end-to-end latency and management for mobile applications that requires multiple computing resources. Such applications are often handled in a fashion of service function chain (SFC), which designates a sequence of virtual
more » ... k functions (VNF) for users' traffic to traverse in order to realize their network application. In order to provide the user a tolerated perceived latency for a SFC-based application, many existing works have taken aim at optimal system-wide placement for SFC in heterogeneous scenarios yet fewer works have studied user-managed placement. In this paper, we formulate the user-managed SFC placement in MEC as a contextual combinatorial multi-arm bandit (C2MAB) problem and proposed BandEdge, a bandit-based algorithm for online SFC placement on edge, which consider user's mobility and service preference while jointly optimizing their perceived latency and service migration delay, and then propose an offline exact approach for the role of performance benchmark. To fit the SFC placement problem in a bandit framework, we model the nodes and links to be arms by viewing them as delays and selects them according to a strategy that balances exploration and exploitation. Finally, we evaluate the proposed algorithm in extensive simulation and Mininet-WiFi emulation experiments, numeric simulation results show that the proposed algorithm can achieve close-to-optimum performance and outperform the greedy learning algorithms by at least 50 percent in terms of scalability. We further validate the superior performance of our proposed method in Mininet-WiFi emulation under different environmental parameters.
doi:10.11575/prism/38760 fatcat:bj4bcnepsfhcvlcvdqd6mdxs6a