The fifth generation (5G) mobile communications system is envisioned to serve various mission-critical services such as industrial automation, cloud robotics, and safety-critical vehicular communications. To satisfy the stringent end-to-end latency requirement of these services, fog computing has been regarded as a promising technology to be integrated into 5G networks, in which computing, storage, and network functions are provisioned close to end users, thus significantly reducing the latency caused in transport networks. However, in the context of fog-computing-enabled 5G networks, the high mobility feature of users brings critical challenges to satisfy the stringent quality of service requirements. To address this issue, service migration, which transmits the associated services from the current fog server to the target one to follow the users’ travel trace and keep the service continuity, has been considered. However, service migration cannot always be completed immediately and may lead to a situation where users experience a loss of service access. In this regard, low-latency service migration plays a key role to reduce the negative effects on services being migrated. In this paper, the factors that affect the performance of service migration are analyzed. To enable low-latency service migration, three main enabling technologies are reviewed, including migration strategies, low-latency, and high-capacity mobile backhaul network design, and adaptive resource allocation. Based on a summary of the reviewed technologies, we conclude that dynamic resource allocation is the worthiest one to research. Therefore, we carry out a use case, where reinforcement learning (RL) is adopted for autonomous bandwidth allocation in support of low-latency service migration in a dynamic traffic environment and evaluate its performance compared to two benchmarks. The simulation demonstrates that the RL-based algorithm is able to self-adapt to a dynamic traffic environment and gets converged performance, which has an obviously smaller impact on non-migration traffic than the two benchmarks while keeping the migration success probability high. Meanwhile, unlike the benchmarks, the RL-based method shows performance fluctuations before getting converged, which may cause unstable system performance. It calls for future research on advanced smart policies that can get convergence quickly, particularly for handling the migration of latency-sensitive services in 5G transport networks.

中文翻译：

---

在5G传输网络中实现低延迟服务迁移的技术[已邀请]

预计第五代（5G）移动通信系统将服务于各种关键任务服务，例如工业自动化，云机器人和安全关键的车辆通信。为了满足这些服务的严格的端到端延迟要求，雾计算已被视为一种有前途的技术，可以集成到5G网络中，在该技术中，计算，存储和网络功能在最终用户附近提供，从而大大减少了传输网络中引起的延迟。但是，在启用雾计算的5G网络的背景下，用户的高移动性功能带来了严峻的挑战，以满足严格的服务质量要求。为了解决这个问题，服务迁移 已考虑将相关服务从当前雾服务器传输到目标服务器，以跟踪用户的旅行轨迹并保持服务连续性。但是，服务迁移不能总是立即完成，并且可能导致用户体验服务访问丢失的情况。在这方面，低延迟服务迁移在减少对正在迁移的服务的负面影响方面起着关键作用。本文分析了影响服务迁移性能的因素。为了实现低延迟服务迁移，对三种主要的使能技术进行了审查，包括迁移策略，低延迟和高容量移动回程网络设计以及自适应资源分配。根据已审查技术的摘要，我们得出结论，动态资源分配是最值得研究的资源。因此，我们执行一个用例，其中在动态流量环境中采用强化学习（RL）进行自主带宽分配以支持低延迟服务迁移，并与两个基准进行比较来评估其性能。仿真表明，基于RL的算法能够自适应动态流量环境并获得融合性能，这对非迁移流量的影响明显小于两个基准，同时保持了较高的迁移成功概率。同时，与基准测试不同，基于RL的方法在收敛之前会显示性能波动，这可能会导致系统性能不稳定。它要求对可迅速收敛的高级智能策略进行进一步的研究，