TY - JOUR TI - Mobile edge cloud architecture for future low-latency applications DO - https://doi.org/doi:10.7282/t3-t53s-b846 PY - 2020 AB - This thesis presents the architecture, design, and evaluation of the mobile edge cloud (MEC) system aimed at supporting future low-latency applications. Mobile edge clouds have emerged as a solution for providing low latency services in future generations (5G and beyond) of mobile networks, which are expected to support a variety of real-time applications such as AR/VR (Augmented/Virtual Reality), autonomous vehicles and robotics. Conventional cloud computing implemented at distant large-scale data centers incurs irreducible propagation delays of the order of 50-100ms or more that may be acceptable for current applications but may not be able to support emerging real-time needs. Edge clouds considered here promise to meet the stringent latency requirements of emerging classes of real-time applications by bringing compute, storage, and networking resources closer to user devices. However, edge clouds are intrinsically local and have a smaller scale and are thus subject to significantly larger fluctuations in offered traffic due to factors such as correlated events and user mobility. In addition, edge computing systems by definition are distributed across multiple edge networks and hence are associated with considerable heterogeneity in bandwidth and compute resources. Considering these challenges, this thesis analyzes the requirements posed by the edge clouds and proposes specific techniques for control, network routing, data migration, and dynamic resource assignment which can be employed to support low-latency applications. The thesis starts by analyzing system-level edge cloud requirements for low-latency by deploying a set of sample AR applications, namely, annotation-based assistance, and smart navigation. A city-scale MEC system is analyzed for achievable latency when running AR applications using existing core clouds as well as the proposed distributed edge cloud infrastructure. Performance evaluation results are presented to understand the trade-offs in key system parameters such as core cloud latency and inter-edge or core-to-edge network bandwidth. The results show that while the core cloud-only system outperforms the edge-only system having low inter-edge bandwidth, a distributed edge cloud selection scheme can approach global optimal assignment when the edge has sufficient compute resources and high inter-edge bandwidth. Adding capacity to an existing edge cloud system without increasing the inter-edge bandwidth contributes to network-wide congestion and can reduce system capacity. Next, a specific network-assisted cloud resource management technique is described that uses the concept of named-object architecture to create a named-object based virtual network (NOVN) inherently supporting application specific routing specifically designed to enable Quality of Service (QoS) in MEC. The results validate the feasibility of the named-object approach, showing minimal VN processing, control overhead, and latency. The results also validate application specific routing (ASR) functionality for an example latency constrained edge cloud service scenario. Further, user mobility and edge cloud system load balancing are handled by enabling dynamic service migration. Container migration is emerging as a potential solution that enables dynamic resource migration in virtualized networks and mobile edge cloud (MEC) systems. The orchestrated, lightweight container migration model is designed and evaluated for a real-time application (license plate recognition) using performance metrics such as the average system response time and the migration cost for different combinations of load, compute resources, inter-edge cloud bandwidth, network, and application latency. The concept of NOVN and service migration are then applied to the advanced driver assistance systems (ADAS) geared towards autonomous driving using Augmented Reality (AR). The experiments show that the low-latency ADAS applications with an average system latency of less than 100 ms for the applications can be supported. The key observations from this study are: (1) machine type plays a crucial role in deciding migration, (2) applications requiring higher computation capabilities, for instance, annotation-based assistance should be offloaded to the closest available lightly-loaded edge cloud, (3) the latency of applications requiring pre-fetched data has fewer avenues for optimization, and (4) service migration should consider network bandwidth, system load, and compute capability of the source and the destination. The work on edge cloud resource assignment and networking motivated the design of a general-purpose control plane that supports the exchange of essential control information (such as compute and network capabilities, current workloads, bandwidth/latency, etc.) between edge cloud domains in a region. The existence of such a control plane enables distributed resource management, application-aware routing, and task assignment algorithms without the requirement of a single point of control. Therefore, the final part of this thesis focuses on creating a lightweight control protocol that can provide neighboring edge clouds with visibility of their computing and network resources along with current load metrics. The proposed design promotes regional awareness of available resources in a heterogeneous multi-tenant environment to enable cooperative techniques such as cluster computing, compute offloading, or service chaining. The design of a specific control plane protocol followed by a system-level evaluation of the performance associated with task assignment and routing algorithms enabled by the framework is presented. The evaluation is based on a prototype system with a heterogeneous network of compute clusters participating in the control plane protocol and executing specified resource sharing algorithms. An application-level evaluation of latency vs. offered load is also carried out for an example time-critical application (image analysis for traffic lane detection) running on the ORBIT testbed confirming that significant performance gains can be achieved through cooperation at the cost of modest complexity and overhead. KW - 5G KW - Cloud computing KW - Electrical and Computer Engineering LA - English ER -