Dynamic application placement in the Mobile Cloud Network

Highlights

• Proposes a holistic cost-optimal management approach for the forthcoming Mobile Cloud Network.

• The proposed method shows significant improvements in aggregate system cost and resource utilisation skewness, and proves to be very capable at accommodating user mobility.

• The evaluations reveal that high cost of user mobility and the amount of coordinated effort is required to accommodate it.

• The proposed method provides an upper bound to which we can contrast more tractable distributed solutions.

Abstract

To meet the challenges of consistent performance, low communication latency, and a high degree of user mobility, cloud and Telecom infrastructure vendors and operators foresee a Mobile Cloud Network that incorporates public cloud infrastructures with cloud augmented Telecom nodes in forthcoming mobile access networks. A Mobile Cloud Network is composed of distributed cost- and capacity-heterogeneous resources that host applications that in turn are subject to a spatially and quantitatively rapidly changing demand. Such an infrastructure requires a holistic management approach that ensures that the resident applications’ performance requirements are met while sustainably supported by the underlying infrastructure. The contribution of this paper is three-fold. Firstly, this paper contributes with a model that captures the cost- and capacity-heterogeneity of a Mobile Cloud Network infrastructure. The model bridges the Mobile Edge Computing and Distributed Cloud paradigms by modelling multiple tiers of resources across the network and serves not just mobile devices but any client beyond and within the network. A set of resource management challenges is presented based on this model. Secondly, an algorithm that holistically and optimally solves these challenges is proposed. The algorithm is formulated as an application placement method that incorporates aspects of network link capacity, desired user latency and user mobility, as well as data centre resource utilisation and server provisioning costs. Thirdly, to address scalability, a tractable locally optimal algorithm is presented. The evaluation demonstrates that the placement algorithm significantly improves latency, resource utilisation skewness while minimising the operational cost of the system. Additionally, the proposed model and evaluation method demonstrate the viability of dynamic resource management of the Mobile Cloud Network and the need for accommodating rapidly mobile demand in a holistic manner.

Introduction

Contemporary cloud services are predominantly hosted by a handful of Data Centres (DCs) on each continent  [1], with no specific regard to their subscribers’ whereabouts.

Furthermore, cloud services are at an increasing rate accessed from Mobile Devices (MDs)  [2] and employed to orchestrate the infrastructure and devices that constitute the Internet of Everything (IoE)  [3]. Collectively, they produce and consume an increasing amount of content, consuming a larger portion of network traffic. Due to intermittent cloud service availability and heterogeneous latency between an MD and a DC, MDs often resort to executing applications locally, as opposed to in the cloud  [4].

The current, centralised infrastructure paradigm, geographically separates users and infrastructure, and thus does not accommodate mission critical, latency sensitive application, and incurs a significant load on the network. The Mobile Cloud Network (MCN)  [5] has been proposed to mitigate these performance inhibitors. In the MCN paradigm, DCs of heterogeneous scale and cost are dispersing throughout the core- and access-networks as a compliment to the existing, centralised, public cloud DCs  [6], see Fig. 1. The proposed infrastructure will span across entire national networks, encompass thousands of subscribers, nodes, and applications. The compute capacity, or DCs, will proposedly reside with a varying degree of proximity to the end-user, from Radio Base Stations (RBSs) to regional hubs. Such an infrastructure will be both cost- and capacity-heterogeneous as a function of depth. Furthermore, the proposed use case of the MCN ranges from hosting virtualised Telecom service  [7], offloading mobile application  [8], mobile edge computing  [9], [10], to orchestrating and enabling Internet of Things (IoT) services  [11], [12].

With increased proximity between end-users and cloud infrastructure, application developers and device manufacturers are now able to realise MD offloading and large Wireless Sensor Network (WSN). As a consequence, they are for example faced with the challenge of determining when and what to offload  [13], given capacity and energy objectives  [14], and how to design their sensor infrastructure  [15], [16], [17], [18].

Furthermore, regardless of how application owners and developers take advantage of the MCN, to realise an MCN, the operator of an MCN shall attempt to meet all resident applications’ performance goals and to manage the operational expenditure of its infrastructure. In such a large, distributed, cost-, proximity-, and capacity-heterogeneous infrastructure, with large set of highly mobile users, the placement of the resident applications and services is an MCN operator’s foremost degree of freedom. The operator of an MCN thus pursues its objectives by actively evaluating the placement of the resident applications in relation to each other, their mobile users, and cost and capacity of the infrastructure’s resources. The fundamental problem addressed in the paper is how to scalably and autonomously manage the infrastructures in such a manner, over a large number of resources, to guarantee application performance, while minimising the incurred cost on the infrastructure given a time-variant demand and a heterogeneous infrastructure, and mitigating resource skewness.

If the resident applications’ placements are not evaluated at the rate of which the applications’ demand is changing, the system cannot guarantee the applications’ performance goals or Service Level Agreements (SLAs), or minimum operational cost or system wide resource usage. The task of evaluating the placement of thousands of applications serving an even greater number of users across hundreds of heterogeneous nodes cannot be done manually, in real-time, by a system administrator. Furthermore, the breadth and depth, and heterogeneity of the infrastructure, and the rate of change in application demand make judging the trade-off, over time, non-trivial, and need to be performed holistically and autonomously by the system.

Restoring to naïve methods, such as placing applications randomly or greedy placing applications as close as possible to the end-user, fails to take into account the time-variant input to the system and the cost- and capacity-heterogeneity of the system, and can thus not guarantee that the applications meet their SLAs, nor does it guarantee that the cost of the system is maintained, and that the load is relatively balanced across the system.

As the MCN paradigm is novel, no implementation exists nor has a full definition materialised. As such, this problem has to the best of our knowledge not been addressed previously. Related performance centric problems can be found in literature for routing, intra- and inter-DC application placement, and in optimal content distribution in Content Delivery Networks (CDNs). However, to the best of our knowledge, none of the presented approaches simultaneously and holistically take into account rapid user mobility and vast resource cost- and capacity-heterogeneous infrastructures.

The contribution of this paper is threefold. (1) A performance model is presented of the proposed MCN infrastructure paradigm. (2) An optimisation problem based on aforementioned objectives is formulated and solved. The optimal solution achieves a 25% reduction in cost compared to the naïve methods. (3) A traceable, iterative, local-optimal algorithm is presented. This algorithm achieves a cost within 5% of the optimal, at only a fraction of the computational overhead.

This paper has the following structure: Section  2 highlights the resource management challenges facing the MCN. Section  3 describes mathematical models associated with the MCN. Section  4 formulates the problems as an optimisation problem. Section  5 describes application placement methods based on the objective function described in the earlier section. Sections  6 Evaluation model, 7 Experiments ​present how the proposed algorithm and models are evaluated by detailing the evaluation model and the experiments, respectively. Section  9 covers related work in tangent research fields. Section  10 summarises this paper’s contributions, conclusion of the experiments and future work.