Smart routing: Fine-grained stall management of video streams in mobile core networks

Abstract

Video traffic has dominated the global mobile data traffic and creates the fundamental need for continuous enhancement and fast evolution in mobile networks so as to accommodate its unprecedented growth. Despite the surging interests in radio access networks (RANs), the latest technologies on dense and heterogeneous wireless networks are shifting the bottleneck of mobile networks to the core networks. However, managing the stalls of video streams in mobile core networks remains challenging. In an evolving mobile system, the core network needs to (i) determine the data rate for each video streaming request, (ii) distribute the video request among multiple sources, and (iii) route the so-generated peer-to-peer flows. In this paper, we exploit user context and propose an optimized routing scheme (termed as smart routing) for stall management in mobile core networks, which adaptively schedules data rates with respect to user context and strategically routes so-scheduled video demands. The proposed smart routing scheme simultaneously addresses the above three aspects by formulating them in a joint optimization problem and solving the formulated problem with a fast algorithm with provable approximation guarantee. Computer simulations validate the efficiency of the proposed scheme.

Introduction

The rapid proliferation of smart phones and tablets is profoundly changing the way we behave and consume content, shifting more and more data traffic from fixed networks to mobile networks. Global mobile data traffic is expected to explode by nearly 11-fold between 2013 and 2018, predominantly fueled by video traffic [1]. Up from 53% in 2013, mobile video will represent 69% of global mobile data traffic by 2018. Such substantial growing demands in both traffic volume and speed are out-pacing the ability of current mobile networks [2], e.g., the long term evolution (LTE) systems of the Third Generation Partnership Project (3GPP), which is designed as “interconnecting middle-boxes” between public networks and mobile customers. On the other hand, recent developments in dense and heterogeneous wireless networks are shifting the bottleneck of mobile systems from radio access networks (RANs) to the backhauls, i.e., mobile core networks [3], [4].

To accommodate the surging demands, components and protocols of LTE core networks are enhanced or re-factored by integrating transparent video caching and incorporating more efficient request-scheduling algorithms [5], [6], [7], [8]. Meanwhile, mobile systems are undergoing fast evolution from the current fourth generation (4G) technologies, i.e., LTE/LTE-Advanced, to the fifth generation (5G) [4] with virtualized networking functions [9], [10], [11]. Various video sources, such as in-system caching [7], mobile content distribution networks (MCDNs) [12], cloud service points [10] and improved routers [13], will be deployed within mobile core networks so as to move data close to customers. Transport and routing technologies, together with computing and storage resources, will be embedded in mobile core networks to build up a converged infrastructure which orchestrates the delivery of data traffic. To promote manageability and flexibility, the control plane is separated from the data plane in LTE systems and evolving mobile systems. As such, control information is delivered in the control plane while data traffic is transferred in the data plane. This enables centralized networking management in mobile core networks, e.g., through a coordinating component implemented at the packet data network (PDN) gateway of an LTE system or the central controller deployed in a software defined networking (SDN)-enabled mobile core [14].

Considering video services, it is critical and challenging to enhance the quality of experience (QoE) of mobile customers. However, the evolving systems still rely on conventional best-effort routing protocols [15], [16] running in lower layers, which fail to make use of the new functionalities to improve QoE and will inevitably let the network bottleneck incline to the core networks. Therefore, mobile core systems are in an urge need of advancing the routing approach with the new network developments and supporting more efficient data delivery and improved QoE.

As we know, QoE is a comprehensive assessment of user-perceived service quality and it depends on a variety of key performance indicators such as latency, data rate, video resolution (spatial and temporal), packet loss and stalling. In particular, stalling is very detrimental to the QoE of video streams [17]. Hence, in this paper, we specifically focus on stall management and aim to provide fine-grained traffic control for video streams in an evolving mobile core network as depicted in Fig. 1. The mobile core consists of a central controller and a set of nodes, including (virtualized) internal content sources,¹ intermediate forwarding servers and edge servers. Edge servers sitting at the edge of the mobile core provide network service to RANs. The solid lines therein represent the data plane, while the dashed lines represent the control plane. In the mobile core network, the central controller periodically gathers network and traffic information, and accordingly shapes network flows and optimizes the network performance.

This architecture is an instant abstraction of future mobile networks with SDN-enabled mobile cores [14]. By mapping the controller to the PDN gateway, edge servers to serving gateways, and intermediate servers to routers, we note that this architecture can also be considered as an enhanced LTE system with built-in content storage. Furthermore, in some cases, internal sources and serving functionalities can be incorporated in evolved routers. This maps the mobile core in Fig. 1 to an information-centric network [13]. Therefore, this model under study is quite general.

Video streaming requests are initiated by mobile devices and enter the core network through edge servers. Existing works on video streaming support in RANs often assume that the required data are always “pre-fetched” at edge servers [18]. Despite the numerous factors that affect user-perceived service quality, managing the stalls that occur during video sessions is essential to achieve better user experience. As the backhaul is becoming the network bottleneck, efficient request routing in the core network is of vital importance for stall management. Request routing in the core network is to determine how to deliver requested video streams to corresponding edge servers. As each video request can potentially be fulfilled by multiple sources through multiple paths inside the network, request routing hereby makes decisions on both source redirection, i.e., how to redirect requests to respective sources, and flow routing, i.e., how to route so-generated peer-to-peer flows in the network.

Current mainstream mobile devices equipped with advanced hardware generally have plenty of memory for buffering video streams. For a particular mobile device, we specifically consider the user context with respect to the video data buffered at user devices as well as the fashion in which the corresponding user consumes such data. With buffer, temporary transmission failure does not necessarily cause stalls (also termed as jitters in [19]). As shown in Fig. 1, transmission curves represent how edge servers deliver data, while playout curves indicate how the data are actually consumed.

Detailed information of these curves is shown in Fig. 2, where x-axis is the time and y-axis is the aggregate amount of data received or consumed. The playback curve is a characteristic of corresponding video and is independent of underlying transmission and user context. Stalls happen only when the buffered data are not able to support normal playback. Depending on the data transmission and the stall-recovery scheme, a stall lasts for certain duration, referred to as stall delay. The term playout lead is adopted from [17], and represents the duration of time that the video can be played using only the data already buffered in the mobile device.

The playout lead plays a critical role in determining whether video playback stalls, and thus is a key factor to improve the QoE of mobile customers. For an interval considered, the playout lead of a video stream at the end of the duration depends not only on the amount of buffered data at the beginning of the interval, but also on the data rate adopted during the interval. Therefore, the data rates of video requests should be adapted according to respective user context, which is the third essential aspect of smart routing, i.e., data-rate selection, in addition to source redirection and flow routing.

In this paper, we focus on stall management of video streams in evolving mobile core networks with virtualized sources. Video requests are collected from edge servers, while adaptive streams are also delivered to edge servers as responses. Here, we consider (YouTube-like) on-demand video services, for which video data are delivered and buffered at mobile devices before playback. Also, it is assumed that stalls occur only when the buffered video data cannot support the playback. Data transmission from edge servers to RANs and within RANs is not considered in this paper, but it is also an active research area in recent years, e.g., heterogeneous wireless access [20] and SDN RANs [9]. We consider video streams that are delivered in a progressive downloading mode, i.e., a mobile device could buffer up to the entire video clip in its local memory. We also assume that the data consuming information at the buffer of remote devices can be efficiently tracked or estimated by core networks and is readily accessible to the controller [17] in the system under study. For example, in an LTE system, the amount of data delivered to remote devices is gathered by the policy and charging rules function (PCRF) component.

In this work, we not only aim to optimize request routing upon the instant request pattern and networking conditions, but also attempt to adjust instant demands of video requests upon user context. To address this, we formulate a joint optimization problem of request routing and data-rate selection. We then conduct extensive analysis on the formulated problem and propose fully polynomial-time approximation schemes to solve the problem for the scenarios with continuous data rates and discrete data rates, respectively. Routing decisions are displayed in a path-flow form, which are readily converted to a routing protocol and dispensed to routing nodes using the techniques from [10], [21].

In summary, our contributions are three-fold. First, to the best of our knowledge, we are the first to exploit user context for fine-grained stall management of video streams in mobile core networks. While most of recent research interests focus on RANs, this paper studies the essential routing problem in core networks, which will be the new bottleneck of mobile systems. As one of the key features of future mobile networks [4], a route of adapting network behavior to user context is sketched out in this work. The problem and the system model we consider here are essential to a robust and high-performance core system in various mobile networks evolving to 5G.

Second, we formulate an optimization framework which jointly considers data rate adaptation, source redirection and flow routing. We aim to manage stalls of video streams by maximizing minimum playout lead over the whole network while maintaining the maximum link utilization under a given threshold. By tracking the average playout lead over concurrent requests, our framework reveals what performance can be achieved by strategic routing, and provides important insights on how the network should be re-factored and whether hardware upgrade is required.

Third, we analyze the hardness of the formulated problem and propose fast algorithms to solve it. The problem is proven to be NP-hard when data rates are discrete, while the resulting linear programming (LP) problem with continuous data rates is shown to be over-sized for LP solvers. Algorithms are developed for both cases and are extensively studied via theoretical analysis and computer simulations. We also show that the proposed algorithms achieve provably approximation ratios with tolerable computing cost and negligible traffic overhead.

The rest of this paper is organized as follows. Section 2 presents related works. In Section 3, we describe the system model, propose the optimization framework and analyze the hardness of the formulated problems. Section 4 gives our solutions to the formulated problems. Simulation studies are presented in Section 5. Finally, Section 6 concludes the paper.