SDN and NFV Integration in Generalized Mobile Network Architecture

20 downloads 28619 Views 540KB Size Report
MOBILE NETWORKS. Any new technology has to fulfil a set of basic assumptions ... network monitoring but on the other and it also provides means for implementing ..... [10] SDN Problem Statement (http://tools.ietf.org/html/draft-nadeau-sdn-.
SDN and NFV Integration in Generalized Mobile Network Architecture Jose Costa-Requena*, Jesús Llorente Santos, Vicent Ferrer Guasch, Kimmo Ahokas, Gopika Premsankar, Sakari Luukkainen Aalto University Espoo, Finland

Ijaz Ahmad, Madhusanka Liyanage, Mika Ylianttila CWC, University of Oulu Oulu, Finland

Oscar López Pérez, Mikel Uriarte Itzazelaia

Edgardo Montes de Oca

Nextel S.A. Zamudio, Spain

Montimage Paris, France

Abstract—The main drivers for the mobile core network evolution is to serve the future challenges and set the way to 5G networks with need for high capacity and low latency. Different technologies such as Network Functions Virtualization (NFV) and Software Defined Networking (SDN) are being considered to address the future needs of 5G networks. However, future applications such as Internet of Things (IoT), video services and others still unveiled will have different requirements, which emphasize the need for the dynamic scalability of the network functionality. The means for efficient network resource operability seems to be even more important than the future network element costs. This paper provides the analysis of different technologies such as SDN and NFV that offer different architectural options to address the needs of 5G networks. The options under consideration in this paper may differ mainly in the extent of what SDN principles are applied to mobile specific functions or to transport network functions only. Keywords— SDN; NFV; 5G, Virtualization

I.

INTRODUCTION

The main driver for the evolution of the mobile Evolved Packet Core (EPC) networks is to pave the way to 5G networks that will require substantially higher capacity, lower latency and massive network access. Furthermore, there will be multiple types of applications with very different requirements, which emphasize the need for dynamic scalability of the network functionalities. The means for the efficient network resource operability seems to be even more important than the future network element costs. Mobile network operators are facing a growing challenge thanks to the explosive increase in data traffic due to the prevalence of smartphones and streamed audio and video services. In this new paradigm, the operators need to manage the traffic load, and meet rising consumer and enterprise expectations for excellent performance while providing ubiquitous broadband connectivity. Operators must also roll out new services and applications rapidly to maintain a competitive edge. Slow service rollouts are no longer *Corresponding author. Email: [email protected]

acceptable. Finally, in every competitive market there is constant pressure to become more efficient; in other words, to maintain or improve performance at a lower operational cost. Existing mobile networks struggle with limitations such as stationary and expensive equipment, complex control protocols, and heterogeneous configuration interfaces. The main goal of this work is to study and apply SDN principles within the mobile networking environments namely SDMN (Software Define Mobile Networks) to be able to address these current limitations. Cloud computing and Network Function Virtualization (NFV) are evolving from the typical IT data center applications to the new areas. Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services). Network virtualization general goals are the aggregation of distributed resources for the common goal and utilizing a shared pool of configurable computing hardware resources for on-demand network access. Cloud computing enables the ability to host network functions such as resource, policy, mobility, security and traffic management, and monitoring within the cloud. In computation and storage environments there are several emerging technologies and enablers that could bring feasible processing potential to be utilized for the mobile network networks applications. The potential use cases for mobile networks include virtual operators, network sharing (overlays) principles and core network element user plane functionality cloudification for the network control overlay functions. SDN decouples control and data planes leveraging standard protocols enabling remote management and operation of data planes to third-party elements. A synchronization protocol is required for communicating both planes; one such protocol is OpenFlow [1]. The benefits of SDN seem quite obvious in the area of cloud computing networking; however the application to the mobile paradigm requires further study.

This paper describes the proposed architecture based on SDN and NFV and its deployment in a testbed to verify that the basic assumptions for 5G networks are fulfilled. The rest of the paper is structured as follows. Section II defines the requirements that have to be addressed when integrating SDN and NFV in mobile networks. Section III describes the NFV and its benefits in addition to the current work done in mobile industry towards its adoption. Section IV describes the current work on SDN integration in mobile networks. Section V presents the proposed architecture that including SDN and NFV. Section VI presents the testbed where the proposed architecture is deployed and used as ETSI Proof of Concept (POC) [2]. In section VII the conclusions are presented. II.

REQUIREMENTS FOR NFV AND SDN INTEGRATION IN MOBILE NETWORKS

Any new technology has to fulfil a set of basic assumptions that facilitate the deployment and adoption. Following are listed some basic assumptions and how they might be applicable to SDN and virtualization technologies proposed for 5G networks. These basic assumptions consist of enhanced resiliency, improved performance, lower latency, seamless migration and inbuilt monitoring. Besides these technical assumptions, the proposed technology should ensure proper service provisioning with a relevant cost reduction to operators. However, there are technical challenges; current mobile networks have around 99.999% availability rate, which needs to be maintained or yet improved to motivate the change [3]. A technology to be part of 5G should provide a clear migration path with proper compatibility with the legacy systems. The integration of SDN and network function virtualization as proposed technologies for 5G should minimize the changes in network elements, thus providing a seamless migration based on operator needs. This allows the incremental updates of network elements in certain parts of the network while keeping legacy elements in other parts of the network.

comprise both a distributed (SDN / NFV-based) QoS measurement system and a centralized evaluation system. Service provisioning and optimization is another assumption required to ensure resource availability. This can be done by a single orchestrator entity. This assumption can be deployed in SDN networks by using control applications that have full view of network configuration. This together with status information provided by network monitoring and data collection systems enable mobile network orchestrator application to optimize service (e.g. latency) and/or resource usage easier than traditional networks that need to rely on signaling. The orchestrator can control multiple network elements via the control applications, potentially from multiple vendors. This enables to introduce new services by writing or modifying the orchestrator whereas in traditional networks all the equipment need to be upgraded to support the new service type. It is assumed that 5G networks will not be implemented following a clean-slate approach; instead, legacy and new SDN based 5G solutions will need to coexist over time. However, in order to exploit the potential of SDN, cooperation between both of these solutions is required, e.g. by introducing abstraction and automation layer for the legacy network part. Above all, cost reduction is a major requirement. Virtualization of the LTE network is expected to contribute in this regard, benefiting from standardized network elements and better resource utilization with SDN. However, virtualized network elements may increase the need for more computing power, more complex network management, and create more complex value networks. The net benefit of SDN in LTE networks should be examined. III.

VIRTUALIZATION AND CLOUDIFICATION IN MOBILE NETWORKS

Security is of utmost importance for 5G networks, as it should be considered for all layers, in network functions as well as physical and virtual elements. Starting with the SDN controller that has access to the whole network architecture, to the actual nodes that perform network functions, the system must guarantee a critical level of security and high availability.

Virtualization of network elements is one of the major technologies proposed for adoption in 5G networks. Virtualization decouples a system’s service model from its physical realization in order to increase the computation performance of a system (e.g., to serve an increased amount of users). The main advantage of virtualization is seen in cloud computing in which technologically distinct systems are brought together onto a single virtualized domain (a collection of physical servers). Different services could be deployed on top of the virtualized system to achieve higher degree of service availability and flexibility.

Another important assumption for an effective deployment is having proper network monitoring. Besides assessing security, network monitoring facilitates verification and validation of Service Level Agreements (SLA), managing performance (Quality of Service QoS) and user experience (Quality of Experience QoE), troubleshooting, and the assessment of optimizations and use of resources. On the one hand, network virtualization sets new requirements for mobile network monitoring but on the other and it also provides means for implementing advanced network monitoring solutions. NFV/SDN enables the integration of cloud infrastructure that provides higher degrees of freedom regarding the placement of measurement points and flexible control of traffic flows. An advanced and effective QoS monitoring solution should

However, distribution of computation over multiple servers reveals the question of load balancing at two levels: VM scheduling and load distribution. The VM scheduler of the cloud platform should, ideally, distribute the computation evenly between the physical servers, simultaneously keeping the number of required servers as low as possible to serve all clients. This allocation can be later modified by migration of VMs to other servers. The only restricting requirements emerge from fault tolerance: the VMs should be allocated to mitigate the effects of failing hardware. On the other hand, service load balancing shares the work between similar service elements. If the workers reside in the same cloud, the load should be distributed evenly to all of them. However, if they are deployed in different locations (e.g., multiple geographically distributed

data centers), additional requirements, such as perceived service latencies, modify the desired allocation. Cloud computing can be described as distributed computing environment over a network which has the capability to run a program or application on many connected computers at the same time [4].The National Institute of Standards and Technology (NIST) [5] defines cloud computing as: “Cloud computing is a model for enabling ubiquitous, convenient, ondemand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. This cloud model is composed of five essential characteristics, three service models, and four deployment models”. The service models are Software as a Service (SaaS), Platform as a Service (PaaS) and Infrastructure as a Service (IaaS). The cloudification of mobile networks has both drivers and constraints. A driver for this is the constant struggle of network operators to maintain their business profitable. One significant source of expense is the use of dedicated network hardware to provide the required services. To avoid this, service providers reach for virtualization of network services. The virtualized functions in turn would be provided on top of a cloud infrastructure. However, a considerable amount of work remains to be done before the cloud-based approach allows reaching a similar quality of service, such as availability and latencies, that dedicated components can provide today. An important part of the cloudification effort is the selection of underlying technologies that control the computing infrastructure of the cloud. There are a number of factors that affect the decision of selecting the appropriate cloud infrastructure. Firstly, the large amount of computation, together with security concerns, lead to the selection of the private cloud approach. The private approach allows the placement of computing resources inside the network architecture, instead of the public Internet, which leads to reduced latencies and improve the control of the overall infrastructure. Secondly, the most rational option is to provide an infrastructure cloud. Cloud services are categorized into three levels of abstraction: IaaS, PaaS, and SaaS. These levels require different levels of competence from the cloud user. For example, a PaaS user must implement the software but the cloud platform manages the service scaling. On the other hand, an IaaS user is required to handle scaling features and distributed system communication. However, an IaaS cloud allows flexibility in selecting the technologies that are used create the network functions. Furthermore, the PaaS cloud is usually created on top of IaaS, making it necessary to select an IaaS platform first. It is a common practice to rely on open-source platforms to avoid vendor lock-ins or build custom features that expand the platform capabilities. We argue that OpenStack is a viable solution [6] to provide the cloud for mobile networks. The large and diverse community reduces the risk of the project being directed in the wrong direction. Also, the amount of supported hypervisors and networking technologies are positive signs. Moreover, OpenStack is becoming widely used,

from small private clouds to commercial solutions, e.g. Rackspace. In this paper, we review the recent development around OpenStack and provide questions that still need to be answered in the future. NFV is highly complementary to SDN. These topics are mutually beneficial but are not dependent on each other. Network functions can be virtualized and deployed without an SDN being required and vice-versa. Telecom or telco cloud is the cloud adapted for telecoms. The principal idea is to exploit cloud computing as infrastructure for the future mobile network deployment and operation. The target is to extend the cloud technologies and its benefits to the network, which is also the goal of the ETSI NFV Initiative. The telecom cloud is not the same as the IT cloud; the telecom industry’s demanding requirements for five nines availability, scalability, reliability and complex networking must be met, and a supplementary approach is required. IV.

INTEGRATION OF SDN IN MOBILE NETWORKS

There are several papers describing the integration of SDN in mobile networks [7-11]. They propose adding SDN agents in the mobile network elements. SoftRAN [7] proposes a centralized architecture as an alternative to the distributed control plane currently implemented in LTE networks. It abstracts out all the base stations deployed in a geographical area as a virtual big-base station while considering all the physical base stations as just radio elements with minimal control logic. These radio elements are then managed by a logically centralized entity which makes control plane decisions for all the radio elements in the geographical area. We call this logically centralized entity, the controller of the big base station. The controller maintains a global view of the radio access network and provides a framework on which control algorithms can be implemented. CellSDN [8] pushes fine-grained packet classification to the access switches, which can be implemented easily in software (e.g., using Open vSwitch). These access switches apply fine-grained rules, specified by the controller, to map UE (User Equipment) trace to the policy tags and hierarchical addresses. To ensure control-plane scalability, a local agent at the base station caches the service policy for each attached UE. Other work [9-11] defines that each base station has an access switch that performs fine-grained packet classification on trace from UEs. Access switches can be software switches (such as Open vSwitch) that run on commodity server hardware. The server can also run a local agent that caches service policies for attached UEs, to minimize interaction with the central controller. The rest of the cellular core consists of core switches, including a few gateway switches connected to the Internet. These core switches perform multi-dimensional packet classification at high speed, but only for a few thousands or tens of thousands of rules. We assume that the packetprocessing hardware can perform arbitrary wildcard matching at different protocol layers e.g. IP or TCP/IP.

V.

SDN BASED MOBILE ARCHITECTURE

Currently there are only a few in-depth scientific contributions dealing with mobile network architectures that combine the concepts of cloud computing, SDN and NFV. First architecture proposals - especially in the context of Cloud-RAN - include the mapping of the network functions that are required for the integration of mobile networks with SDN technology. These functions are only the mobile network control functions, i.e., MME, HSS, PCRF and the control planes of S/P-GW. Additional functions include transport, load balancing, security, policy, charging, monitoring, QoE or resource optimization. These functions run on the Mobile Network Cloud as SDN applications and enforce the desired function by means of SDN technology. With this approach, the user plane is only composed by strategically located SDN capable switches and regular switches. SDN switches could either replace partly or entirely the current mobile transport network [12-13]. This consolidated architecture is shown in Fig. 1. The required EPC network elements run on the cloud to benefit from virtualization. Latency constrains could affect the deployment location of some compute nodes running virtual. Some strategic functions could be placed close to the eNBs or even on some switches, creating a decentralized cloud. In the proposed architecture, the EPC network elements maintain current 3GPP interfaces to favor migration from legacy mobile networks. That will allow a seamless migration

Fig. 1. SDN based consolidated architecture towards 5G

Fig. 2. Seamless 3-Step Migration Towards SDN Enabled EPS

as previously stated in section I. Fig. 2 represents a 3-step migration scenario using OpenFlow as SDN communication protocol. The first use case (i.e. UC1) follows a traditional routed LTE architecture, with legacy nodes. The second use case (i.e. UC2) introduces the SDN technology for managing layer-2 switched paths on the mobile core network, while still maintaining the legacy nodes. This scenario constitutes a hybrid approach allowing isolation of tenant networks using standard encapsulation technologies, i.e. VLAN or MPLS. Finally, the third use case (i.e. UC3) depicts a fully compliant SDN network. This is because the 3GPP tunneling specifics are not supported by current OpenFlow specifications and therefore have been replaced by compliant and more effective encapsulations. These use cases can coexist allowing hybrid deployments to take advantage of existing network appliances. VI.

TESTBED AND RESULTS

The proposed architecture is deployed in a testbed where the usage of the proposed technologies such as NFV and SDN are analyzed against the basic requirements. The testbed consists of two eNodeB provided by Nokia, OpenFlow enabled MPLS switch provided by Coriant Oy, traffic monitoring probe provided by EXFO, S/P-GW is open source nwEPC (SAE Gateway), Ryu SDN controller and the rest of the components (MME, NAT and CES) have been implemented by the research group at Aalto. The SW components are running on Aalto data center using Openstack Icehouse release. The HW components are blade servers that run the cloud as a separate FlexNIC with Intel Xeon E5-2665 (2.4-3.1 GHz, 64-bit, 8 cores, HyperThreading), 128 GB DDR3-1600 RAM, 150 GB hard disk storage and 10 Gbps HP Flex-10 networking drivers. In the testbed we have implemented the three scenarios described in section V (i.e. UC1, UC2 and UC3) to verify the basic requirement of migration. We demonstrate how to deploy vEPC based on current standard network elements where each of them are running on different NVFs. Each of the different network (NW) functions (i.e. MME, S/P-GW and FW) will be running on their own virtual machines in the cloud. A set of probes are included to support the monitoring requirement and provide continuous information on the network status. The eNodeBs provided by Nokia are installed in Aalto premises with own network connection between the eNodeB and the data center where the rest of the NW functions are running. There is GTP tunneling between the eNB and the S/P-GW in the first scenario (UC1) but TAG component removes the GTP in the mobile backhaul for UC2 and UC3. The TAG maintains QoS using MPLS tagging for identifying the flows in the OpenFlow switches (OFS). As a result, in UC1, the UE data packets are routed between the eNB and the S/P-GW following current 3GPP specifications based on GTP tunnels. In both UC2 and UC3 we use SDN to add layer 2 MPLS tagging to the GTP packets so we can perform traffic engineering in the backbone, in order to address the requirement on service provisioning and optimization. UE data packets are switched from the eNB to the S/P-GW across the core network using several paths. Load balancing between OFS#1 and OFS#2 links is possible based on the MPLS identifiers. In UC3 we demonstrate the cost reduction

requirement where the usage of SDN replaces completely the data plane part of standard NW elements such as S/P-GW. Most of the NW elements are running on the cloud using commodity servers. Besides cost reduction, this scenario shows effective service provisioning and optimization where additional virtualized middle boxes could be added to provide NFV functions for managing specific flows and deploy new services. These middle boxes could deflect HTTP packets to proxy servers for optimal caching [14] or the middle boxes could identify suspicious flows and redirect them to firewalls or honeypots to fulfil the security requirement. The cost saving are obtained not only from the virtualization but also from the simplification of the transport network by removing the GTP tunneling encapsulation and using the eNB for sending the data packets in a specific formatting, which in this case is supported by OpenFlow. The backbone network switches packets based on MPLS/VLAN identifiers leading to better utilization and traffic engineering.

requirements identified as mandatory for the adoption of any new technology in mobile networks. A testbed based on this architecture is submitted as ETSI PoC. This testbed has been implemented using off the shelf eNodeBs and SDN capable MPLS switches. The outcome shows that integration of SDN and NFV addresses some of the needs of 5G mobile networks. The results also demonstrate the benefits of SDN that when used in the backhaul improve the efficient and optimal usage of resources with reduced overhead. However, we have identified certain limitations in the proposed technologies. NFV where all NW elements run on the cloud and rely on virtualization, do not provide the necessary reliability and robustness. The testbed results shows high latency when moving VM with NW elements (e.g., MME or S/P-GW) because of HW failure or when there is need for additional processing resources. Thus, reliability and robustness need to be addressed in the proposed virtualization platform.

We tested the delay of migrating VMs running network elements handling the signaling, i.e. MME and the user data, i.e. S/P-GW. For the performance measure, we captured the packets using Wireshark and transmitted echo requests at 0.01 sec. intervals, then manually instruct the Nova service to live migrate the VM instances. Fig. 3 reveals that a live migration of an MME may introduces delays of up to 2 sec., which may dramatically affect the ongoing connections. We also performed the migration of S/P-GW obtaining similar results.

ACKNOWLEDGMENT This work has been performed in the framework of the CELTIC-Plus project C2012/2-5 SIGMONA. The organizations on the authors list would like to acknowledge the contributions of their colleagues to the project. REFERENCES [1] [2]

[3]

[4]

[5]

Fig. 3. Downtime of live migration VM with MME

With the testbed we also demonstrate the first benefit of using SDN to simplify the transport network after removing GTP from user plane. The additional layer 2 tags added for SDN compatibility are typically supported by network interface cards and do not impact the 1500 bytes standard Ethernet payload size. The results in Table I reveal that GTP encapsulated packets are often fragmented, resulting in an overhead of up to a 50%. TABLE I. Message ACK Data Data_Fragment 1 Data_Fragment 2

PACKET OVERHEAD DUE TO FRAGMENTATION Packet 76 1536 1500 72

Payload(B) 40 1500 1464 36

Header(B) 36 36 36 36

Overhead 52,6% 2,34% 2,4% 50%

VII. CONCLUSIONS The proposed architecture includes the integration of SDN and the usage of NFV as a basis for 5G networks. A novel integration of SDN is proposed in order to address the

[6] [7] [8]

[9] [10] [11] [12] [13]

[14]

OpenFlow specifications (https://www.opennetworking.org/) PoC#26: Virtual EPC with SDN Function in Mobile Backhaul Networks (http://nfvwiki.etsi.org/index.php?title=Virtual_EPC_with_SDN_Functi on_in_Mobile_Backhaul_Networks) Carrier-Grade: Five Nines, the Myth and the Reality (http://www.pipelinepub.com/0407/pdf/Article%204_Carrier%20Grade_ LTC.pdf) Ferguson A, Guha A, Liang C, Fonseca R, Krishnamurthi S,, Hierarchical Policies for Software Defined Networks, Proceedings Hot Topics in SDN, 2012 http://cs.brown.edu /~sk/Publications/Papers/Published/fglfk-hier-pol-sdn/ NIST Definition of Cloud Computting, U.S. Department of Commerce. (http://csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf Companies Supporting the OpenStack Foundation (http://www.openstack.org/foundation/companies/) SDN networks (http://www.stanford.edu/~skatti/pubs/hotsdn13softran.pdf) Software-Defined Cellular Core Networks ( http://www.opennetsummit.org/pdf/2013/research_track/poster_papers/f inal/ons2013-final19.pdf) Taking Control of Cellular Core Networks (http://arxiv.org/pdf/1305.3568v1.pdf ) SDN Problem Statement (http://tools.ietf.org/html/draft-nadeau-sdnproblem-statement-00 ) Toward Software-Defined Cellular Networks (http://www.cs.princeton.edu/~jrex/papers/ewsdn12.pdf ) Costa-Requena J., "SDN integration in LTE mobile backhaul networks", IEEE Information Networking (ICOIN), 2014 Feb 10-12, Thailand. Costa-Requena, J; Kantola, R; Llorente, J; Ferrer, V; Manner, J; Yi Ding, A; Liu, Y; Tarkoma, S, "Software Defined 5G Mobile Backhaul", in 1st International Conference on 5G for Ubiquitous Connectivity", Nov 26-27, 2014 Levi, Finland. Costa-Requena, Jose; Kimmerlin, Maël; Manner, Jukka; Kantola, Raimo, "SDN optimized caching in LTE mobile networks", Information and Communication Technology Convergence (ICTC), 2014 International Conference on , vol., no., pp.128,132, 22-24 Oct. 2014