Distributed Agent-Based Orchestrator Model for Fog Computing

Abstract

Fog computing is an extension of cloud computing that provides computing services closer to user end-devices at the network edge. One of the challenging topics in fog networks is the placement of tasks on fog nodes to obtain the best performance and resource usage. The process of mapping tasks for resource-constrained devices is known as the service or fog application placement problem (SPP, FAPP). The highly dynamic fog infrastructures with mobile user end-devices and constantly changing fog nodes resources (e.g., battery life, security level) require distributed/decentralized service placement (orchestration) algorithms to ensure better resilience, scalability, and optimal real-time performance. However, recently proposed service placement algorithms rarely support user end-device mobility, constantly changing the resource availability of fog nodes and the ability to recover from fog node failures at the same time. In this article, we propose a distributed agent-based orchestrator model capable of flexible service provisioning in a dynamic fog computing environment by considering the constraints on the central processing unit (CPU), memory, battery level, and security level of fog nodes. Distributing the decision-making to multiple orchestrator fog nodes instead of relying on the mapping of a single central entity helps to spread the load and increase scalability and, most importantly, resilience. The prototype system based on the proposed orchestrator model was implemented and tested with real hardware. The results show that the proposed model is efficient in terms of response latency and computational overhead, which are minimal compared to the placement algorithm itself. The research confirms that the proposed orchestrator approach is suitable for various fog network applications when scalability, mobility, and fault tolerance must be guaranteed.

Keywords: agent-based orchestrator; distributed orchestrator; fog computing; fog service orchestration; internet of things; service placement.

Figure 1

General Fog Computing architecture.

Figure 2

Model of the proposed distributed orchestrator architecture.

Figure 3

Structure of the agent-based orchestrator component of one fog node.

Figure 4

Relocation of services in case of low computational resources.

Figure 5

New service deployment, depending on the destination fog node given by the best placement algorithm.

Figure 6

Disconnection from the fog node, when the user end-device moves between fog nodes.

Figure 7

Fog node failure detection using heartbeat messages.

Figure 8

Flow chart of the service placement finding method.

Figure 9

Hardware setup used for the experimental evaluation of the proposed model.

Figure 10

Evaluation of the latency of the new service starting: (a) Absolute latency; (b) Relative latency caused by different parts of the placement algorithm.

Figure 11

Evaluation of service redistribution latency: (a) Absolute latency; (b) Relative latency caused by different parts of the placement algorithm.

Figure 12

Evaluation of the service placement optimization process: (a) Time dependency on the number of particles (200 epochs used); (b) Time dependency on the number of epochs (200 initial particles used).

Figure 13

Scalability of the service placement finding algorithm.