As a newly emerged promising computing paradigm, Multi-access Edge Computing (MEC) is capable of energizing massive Internet-of-Things (IoT) devices around us and novel mobile applications, especially the computing-intensive and latency-sensitive ones. Meanwhile, featured by the rapid development of cloud-native technologies in recent years, delivering Artificial-Intelligence (AI) capabilities in a microservice way in the MEC environments comes true nowadays. However, currently MEC systems are still restricted by the limited computing resources and highly dynamic network topology, which leads to high service deployment/maintenance cost. Therefore, how to cost-effectively and robustly deploy edge AI microservices in failure-prone MEC environments has become a hot issue. In this study, we consider an edge AI microservice that can be implemented by composing multiple Deep Neural Networks (DNN) models, in this way, features of different DNN models are aggregated and the deployment cost can be further reduced while fulfilling the Quality-of-Service (QoS) constraint. We propose a Three-Dimension-Dynamic-Programming-based algorithm (TDDP) to yield cost-effective multi-DNN orchestration and load allocation plans. For the robust deployment of the yield orchestration plan, we also develop a robust microservice instance placement algorithm (TLLB) by considering the three levels of load balance including applications, servers, and DNN models. Experiments based on real-world edge environments have demonstrated that the proposed orchestration and placement methods can achieve lower deployment costs and less QoS loss when faced with edge node failures than traditional approaches.

作为一种新出现的有前途的计算范式，多接入边缘计算 (MEC) 能够为我们周围的大量物联网 (IoT) 设备和新型移动应用程序提供活力，尤其是计算密集型和对延迟敏感的应用程序。同时，随着近几年云原生技术的快速发展，在MEC环境中以微服务的方式提供人工智能（AI）能力正在成为现实。然而，目前 MEC 系统仍然受限于有限的计算资源和高度动态的网络拓扑，导致高昂的服务部署/维护成本。因此，如何在易出故障的 MEC 环境中经济高效、稳健地部署边缘 AI 微服务成为热点问题。在这项研究中，我们考虑可以通过组合多个深度神经网络 (DNN) 模型来实现的边缘 AI 微服务，通过这种方式，聚合不同 DNN 模型的特征，可以进一步降低部署成本，同时满足服务质量 (QoS) ）约束。我们提出了一种基于三维动态编程的算法 (TDDP)，以产生具有成本效益的多 DNN 编排和负载分配计划。为了稳健部署收益编排计划，我们还通过考虑三个级别的负载平衡（包括应用程序、服务器和 DNN 模型）开发了一种稳健的微服务实例放置算法 (TLLB)。