Scalable Mobile Core Network Architecture for All-IP Wireless Access

Scalable Mobile Core Network Architecture for All-IP Wireless Access Masugi Inoue† Khaled Mahmud† Hidetoshi Yokota‡ Takeshi Kubo‡ Akira Idoue‡ †New Generation Mobile Network Project Communications Research Laboratory 3-4 Hikarino-oka, Yokosuka, Kanagawa 239-0847, Japan

‡KDDI R&D Laboratories, Inc. 2-1-15 Ohara, Kamifukuoka Saitama 356-8502, Japan

{inoue, kmahmud}@crl.go.jp

Abstract As a variety of wireless access technologies have emerged and 3G systems are evolving towards “all-IP” where both voice and data services are converged into IP core networks, IP mobility has become increasingly important. IP micro-mobility protocols have been proposed for managing local mobility. They have a hierarchical network structure, which contributes to decreased network overheads by minimizing signaling traffic for registration. In this paper, we first define a mobile core network and introduce traffic models for it. We then point out that micro-mobility protocols based on this hierarchical structure have a network bottleneck in the gateway node due to packet routing that could worsen network scalability. We then propose an architecture for the mobile core network. The architecture has a non-hierarchical flat topology and has no single gateway node routing packets, making it possible to provide fast and low-latency packet forwarding and delivery to mobile nodes.

Keywords IP network, mobile network, micro-mobility, scalability

1. Introduction The demand for better telecommunication services has led to the development of a number of wireless access technologies such as wireless PANs (Bluetooth, ZigBee, UWB), wireless LANs (WiFi, HiperLAN, MMAC), wireless MANs, cellular systems (IEEE802.20, second-, third- and fourth-generation cellular systems). The third-generation cellular systems, which were initially designed to use the legacy non-IP networks of the second-generation systems, are evolving toward all-IP networks where both voice and data services are converged into IP core networks. All these developments have made us certain that IP mobility will become increasingly more important. Mobile IP [1], which has been standardized for this purpose and has already been adopted by commercial networks, e.g., cdma2000, allows a mobile node (MN) to bring its home address when it leaves the home network to go to a foreign network. When the MN is in the foreign network, the packets destined for it are first received by its home agent (HA) before being forwarded to it. MNs do not always populate the home network in large-scale mobile networks and the most packets addressed to them are transferred to them via the HA. Thus, in many cases, the location of the HA is not essential in mobile networks.

{yokota, t-kubo, idoue}@kddilabs.jp Furthermore, as the MN moves while it is communicating, the point of attachment to the network is frequently changed. These cause significant network overheads in terms of increased delay, packet loss, and signaling. Dynamic home agent assignment [2] is one solution to this problem, providing a mechanism of selecting the HA nearest to the MN during initial registration. IP micro-mobility protocols are yet another approach and designed for environments where MNs change their point of attachment to the network so frequently that the base Mobile IP mechanism causes the problems described above. Micro-mobility aims to handle the local movement of MNs without interaction with the Internet supporting Mobile IP. IP micro-mobility contributes to reducing delay and packet loss during handoff and eliminates registration between MNs and HAs when MNs remain inside their local area of coverage. Well-known IP micro-mobility protocols such as Cellular IP [3], Hawaii [4], and Hierarchical Mobile IP [5], all have a hierarchical structure, which is recognized as an efficient topology to minimize signaling for registration. We, however, consider that the hierarchical architecture is not necessarily efficient for packet delivery and has a negative impact on scalability as the number of MNs increases. We propose a new architecture for a large-scale mobile network and show that it can avoid concentrations of registration information on a single node and make the mobile network more scalable. In the following, we first discuss the traffic to be supported by the mobile core network and define the volume for target traffic. We then briefly introduce IP micro-mobility protocols and address problems with their hierarchical network architecture. We then propose a scalable mobile core network architecture and describe its features.

2. Mobile Core Network 2.1 Scope We define a mobile core network as one that has any kind of IP-based radio access point, such as the access points of wireless LANs, connected to it. Several mobile core networks are interconnected to one another through ultra-high-speed optical networks with tera-bit-, or perhaps peta-bit-class speed in the future. The scale of a mobile core network should be determined in terms of the traffic to be handled, the available technologies for transmission and switching, and the deployment scenario from the business point of view. Figure 1 is an outline of a mobile core network.

The smallest mobile core network is considered to cover an entire office building, that for a department store, the entire area of a central station, or an outdoor amusement park. A business complex would be covered by a larger network. A much larger mobile core network or a combination of several small mobile core networks would cover the entire area of a small city or part of a dense metropolitan area such as a business district, an area along a belt of highway, or an arterial railroad.

Correspondent Node (CN)

Home Agent (HA)

The Internet

Mobile Core Network

Mobile Core Network

Mobile Core Network

Access Point (AP)

would cover an area with a million people during business hours, all of whom had cell phones and 10,000 or 1% of the phones were connecting and generating traffic at every instant. Figure 2 outlines this traffic model. If every connected phone generated traffic at 100 kilobits per second (kbps), the total volume of traffic the mobile core network could support is 1 Gbps. One-hundred kbps traffic includes high-quality voice communications and low-quality video streaming. If the variety of forthcoming multimedia applications increases the unit volume of traffic to 1 Mbps, 10 Mbps, and 100 Mbps, the total traffic throughout the mobile core network would increase to 10 Gbps, 100 Gbps and 1 Tbps. A long-term target to be achieved around 2010, on the other hand, is to evolve a mobile core network that covers more-populated areas that generate much more traffic. One example is the center of Tokyo, which is composed of ten wards and has a five times larger daytime population than that of the center ward. Another is a ring-shaped area along a loop road or railroad with a diameter of five to ten kilo meters. Tokyo

Mobile Node (MN)

Japan

Figure 1. Scope of mobile core network. 2.2 Target Traffic Models It is not easy to predict the future volume of mobile Internet traffic to be accommodated by mobile networks because this is still increasing at a rapid pace. In the following, Tokyo is taken as a typical metropolitan area and we try to model the mobile traffic that our proposed mobile core network architecture should be able to support. In the modeling, we use the White Paper [6], which offers a variety of statistical data on telecommunications services throughout Japan in fiscal 2001. We first looked at the statistical data on mobile traffic generated by cell phones throughout the whole of Japan because there is no such data specific to Tokyo. From the data, we found that during the peak hour in mobile communications from 7 to 8 PM, 4.63 billion calls were generated from cell phones and the total connection time for those calls was 125 million hours in 2001 throughout Japan. These data indicate that about 340,000 cell phones were connected in every instant in that hour. Japan has a population of about 127 million, and about 83 million cell phones are used. Therefore, about 63 percent of people have cell phones and about 0.4% of these phones are being connected to the network in every instant in the busiest hour. The metropolis of Tokyo is composed of twenty-three wards. In the following, we focus on a mobile core network that will cover a ward in the center of the metropolis. From a white paper on the metropolis of Tokyo, it was clear that the most populated ward during business hours has a residential and working population of about 800,000 within an area of only ten square kilometers. Using all this data as a reference for modeling future traffic, we overestimated that within a few years a mobile core network

A Mobile Core Network 1 Gbps to 1 Tbps aggregated traffic

100 kbps to 100 Mbps traffic from each active MN

One Percent Active Mobile Nodes

10 Km2 One Million Mobile Nodes

Figure 2. Model of mobile core network covering future typical dense ward in center of Tokyo.

3. Issues in Micro-Mobility 3.1 Micro-Mobility Protocols Localized mobility management (LMM) [7], or so called micro-mobility, is an important approach in all-IP networks to handle local movement of MNs without generating HA traffic [8]. Cellular IP [3], HAWAII [4] and Regional Registration [5] are major protocols providing micro-mobility. The first two employ the host routing approach while the last one is based on the tunneling technique. Thus, although they use different approaches, all of them have a common hierarchical network structure or tree topology to handle local mobility. These three protocols with their hierarchical network structure have an advantage in simplifying upward routing and in minimizing signaling with limited MN movement. However, the volume of a routing table that a routing node has

to maintain increases in a hierarchical structure as the node layers increase. 3.2 Issues in Hierarchical Structure 3. 2. 1 Layer-3 Micro-Mobility Approach Figure 3 outlines a typical hierarchical network structure to provide IP-layer (layer-3) micro-mobility management. The gateway foreign agent (GFA) is the top node and below it is a domain configured of six other nodes. Even if a mobile node changes its point of attachment to the network from foreign agent (FA) 1 to FA2, FA3, or FA4 in this structure, the MN does not need to re-register its new point of attachment with a home agent (HA) somewhere outside this network unless the MN exits this domain. This is why micro-mobility protocols employing a hierarchical network structure can minimize signaling traffic. In this network, FA1, for example, knows that MN-A is being connected with the access network below FA1. The upper node, regional foreign agent (RFA), knows that MN-A is under FA1 and that MN-B is under FA2. In this way, the top node, GFA, knows which direction it should forward packets. More precisely, GFA holds a routing table where pairs of an MN’s IP address and the number of the outgoing port to which GFA forwards packets addressed to the corresponding MN are recorded. When GFA receives an IP packet, it finds the final destination, which is an MN’s IP address, by looking into the header of the IP packet. Then, GFA searches the outgoing port that matches the destination MN’s IP address. When a mobile core network with this tree topology with layer-three micro-mobility capability covers this densely populated area of one million users, it is obvious that the top node has to hold a routing table that registers one million entries.

create a tree structure with the Spanning Tree Protocol. Because of the self-learning function the L2 switches have, they have the capability of learning which direction they should forward a packet to. In Fig. 2, for example, after the learning process finishes, SW1 knows that packets whose destination media access control (MAC) address is “Ma” should be forwarded to SW2. In contrast to the layer-three micro-mobility protocols, with this architecture the gateway (GW) does not need to hold as many routing entries as the number of MNs inside the switched-LAN, since these are directly connected to the same segment. This approach is thereby apparently more scalable than the L3 approach. However, when GW transmits a packet destined for an MN, it first has to resolve the MAC address of the MN and to store the result in its address resolution protocol (ARP) cache. To do this, GW broadcasts an ARP request packet where the IP address of an MN is embedded. When the corresponding MN receives this packet, it returns the ARP reply packet that has the MAC address as well as the IP address inside. After GW has received the ARP reply packet, it adds the pair of these two addresses as a new entry into the ARP cache. In the above, the L2 micro-mobility approach still has a problem with scalability. The reasons are as follows: !

The number of MNs visiting the switched-LAN segment increases the number of ARP entries,

!

Each ARP entry is cleared within a pre-determined time after registration, and

!

Creating a new ARP entry needs a certain amount of time to complete the ARP request and reply process. MG’

Ma

MNa

SW2

ARP Reply

[IPa,Ma]

Node A: FA1 Node B: FA2

RFA1

Node A

Node A

RFA2

FA3 Node B

Node B

MR

[IPb,Mb]

Node C

MNb

IPx

IPa

ARP Request ARP Reply

[MG’]

IP packet ARP cache

GW

IPa IPb

[MG] [MR]

R

Ma Mb

Figure 4. Typical hierarchical network with layer-2 micro-mobility management.

FA4

Node C

SW3

MG

ARP Request

ARP Reply

Node C: FA3 Node D: FA4

FA2

IPa

Switched-LAN SW1

GFA

FA1

IPx

ARP Request

Node A: RFA1 Node B: RFA1 Node C: RFA2 Node D: RFA2

Node D

Node D

Figure 3. Typical hierarchical network with layer-3 micro-mobility management. 3. 2. 2 Layer-2 Micro-Mobility Approach While the above approaches focus on the IP layer, micro-mobility protocols on the link layer can also be considered. Figure 4 shows a typical network using this approach. Layer two (L2) switches in the switched-LAN can

3. 2. 3 Summary In a hierarchical or tree-topological network architecture, the routing table in layer-3 approaches or the ARP cache in layer-2 approaches in the top node increase as the number of MNs increases. This increases the time necessary for searching matched entries. What we have to take into account at the same time is that as the number of MNs increases they generate much more traffic. Thus, the increase in the number of MNs has a significant impact on the size of the routing table or the ARP cache as well as the volume of traffic, which goes through the top node.

Although emerging new technologies such as a high-capacity memory for the table, switching fabric architecture and hardware design, and a high-speed route search algorithm would contribute toward alleviating these bottlenecks, we think that we will need a much more fundamental solution to this problem.

the management network is a logically deployed network for providing authentication, authorization, and accounting (AAA), location management, and name service. One or more gateways (GWs) can be placed in the forwarding network, thus enabling different Internet service providers (ISPs) to connect to the same forwarding network.

4. Scalable Mobile Core Network Architecture

In this architecture, we considered that most users would be covered with one LMM network while they were communicating and only a few high-mobility users traveling from one LMM network to another. This means that most users would be able to enjoy high-speed seamless communications while moving because the forwarding network, which is the essence of an LMM network, provides fast and low-latency packet delivery to mobile nodes. When a user travels to another LMM network, the packets destined for the user are assumed to be forwarded with a macro-mobility protocol such as Mobile IP.

4. 1 Design Principles The considerations above motivated us to propose a novel architecture for a large-scale mobile core network. It is now clear that the highest node in the hierarchy always holds the routing information or MAC addresses for all MNs in the network. To eliminate this bottleneck, we applied the following principles in designing our proposed architecture: !

! !

Routing information must not be concentrated on a single node,

In summary, the following represent the major features of this architecture.

ARP cache must not be maintained by a single node, and The network for user data and that for signaling such as location registration or authentication should be separate.

!

As GWs do not need to route incoming packets, they have no routing information or ARP cache. We believe that this contributes to high scalability.

!

The forwarding network for packet transport and the management network are logically separated to provide high-speed packet forwarding.

!

The forwarding network has a non-hierarchical or flat architecture that cannot produce bottleneck nodes.

4. 2 Proposed Architecture 4. 2. 1 Overview Figure 5 is an overview of the proposed architecture. We can see it focuses on one localized mobility management (LMM) network. We considered that one LMM network should be able to cover a densely populated part of a metropolitan area as was described in Sec. 2.2. A variety of wireless and wired access networks are connected to the LMM network. To cover a larger area, another LMM network should be deployed and connected to one another.

4. 2. 2 Forwarding Network The forwarding network can be constructed with L1 or L2 devices such as WDM networks or switched LANs. The topology of the forwarding network could be ring, star or mesh although it is not explicitly described in the figure. The most important point in the forwarding network is that GWs, which connect the forwarding network with outer networks, do not have routing information or an ARP cache, thus eliminating the

The LMM network is composed of two networks: a forwarding network and a management network. The forwarding network is a physically-deployed non-hierarchical network only designed for high-speed packet transport, while

Localized mobility management (LMM) network

CN

Management network (MNW)

Internet (ISP2) GW

Access network (ANW)

Loc. Svr.

LMA

FWD LUD

visitor list MNa IPa MNb IPb FWD LUD

LMA

home AAA

Name Svr.

FWD LUD

FWD LUD

Forwarding network (FNW) for packet transport

(proxy) AAA

Internet (ISP1) LMA

FWD LUD FWD LUD

MNb MNa

(IPa)

GW

CN

LMA MNc

(IPb)

HA

MNc (IPc)

visitor list IPc

LMA: Localized Mobility Agent FWD: Forwarding LUD: Look-Up and Deliver

Figure 5. Scalable mobile core network architecture.

routing of incoming packets. We believe that this contributes to high scalability. We called the mobility agent the “Localized Mobility Agent” or the LMA according to Ref [7] and it has a two-tier structure. The upper tier of the LMA (“Forwarding”, or FWD) is devoted to packet forwarding and the lower tier (“Look-up and Deliver”, or LUD) extracts packets for the registered MNs. Once a packet is transferred on the forwarding network, it is delivered to all the LMAs without interventions in routing. When each LMA receives a packet, it checks if the destination IP address is registered by referring to the visitor list. If so, the LMA transfers the packet toward the access network. The MAC address of the target MN is obtained via the registration procedure and the LMA only caches those of the accommodated MNs. 4. 2. 3 Procedures The registration and the packet transmission procedures are done as follows. Registration: When the MN visits the access network, it registers the location with the LMA that manages the network. The HA needs to become involved when the MN uses its home address, otherwise the MN may be assigned an IP address that is routable throughout the LMM network. Signaling messages for authentication are transferred to the (proxy) AAA in the management network via a separate path. When the registration is successful, the LMA stores the IP address and the MAC address of the MN in its visitor list. Packet Transmission: When a CN on the global network sends a packet destined for the MN, it is delivered to the GW via the HA or by normal IP routing. The GW then transmits the packet on the forwarding network, and it is received by all LMAs. Note that the GW does not need to know to which LMA the target MN belongs and only the LMA that accommodates the MN transmits the packet toward the access network to which the MN is connected. Also, when the MN sends a packet toward the LMA, it just transmits it on the forwarding network. If the destination address of the packet is on the visitor list of some LMA, the packet is taken into the LMA to be delivered to the MN. Likewise, if the destination address is outside the network, the GW routes it in the appropriate direction.

hierarchical structure need to have such a table with one million entries if the mobile core network has one million MNs. This clarifies the advantages of this architecture. Rings, meshes and stars are possible physical network topologies for arranging the proposed architecture. Of these, the ring topology is the best because it is more reliable and scalable than star and mesh topologies. The star topology needs to have a highly reliable and high performance center node. Thus, the center node could become a bottleneck. The mesh topology, on the other hand, needs more links to connect nodes to one another and each node should have routing capability even though it is more reliable than the star topology. In the deployment of a network, the selection of a topology depends on a number of factors including the scale of the network to be deployed, the necessity for reliability, fault tolerance, and deployment costs.

5. Conclusion In this paper, we first defined a mobile core network and introduced a scenario where the volume of future traffic should be covered by this network. We then pointed out that micro-mobility protocols based on a hierarchical structure could create a bottleneck that worsens network scalability. We proposed a novel architecture for the mobile core network. The forwarding network of this architecture has a de-centralized and two-tier structure, making it possible to provide fast and low-latency packet forwarding and delivery to mobile nodes. Based on this architecture, we developed an experimental network with a ring topology and have been evaluating its performance [9].

References [1] C. Perkins “IP Mobility Support for IPv4,” RFC3344, IETF, Aug. 2002. [2] M. Kulkarni, et al., “Mobile IPv4 Dynamic Home Agent Assignment

Framework,”

draft-kulkarni-mobileip-dynamic-assignment-01, IETF, June 2003. [3] A. Valkó “Cellular IP: A New Approach to Internet Host Mobility” ACM SIGCOMM Computer and Communication Review, Vol. 29, No.1, pp. 50-65, January 1999. [4] R. Ramjee, et al. “HAWAII: A Domain-Based Approach for Supporting Mobility in Wide-area Wireless Networks” Proc. IEEE Int’l Conf. Network Protocols, 1999. [5] E. Gustafsson, et al. “Mobile IPv4 Regional Registration” draft-ietf-mobileip-reg-tunnel-05, IETF, September 2001.

Making the LMM network flat enables the registration information to be de-centralized and there is only one lookup of the visitor list by the LMAs, which contributes greatly to increasing the scalability of the LMM network. Unlike a L2 switch, the LMA can manage the IP and MAC addresses of the MNs, which removes the burden of MAC address management in the GW.

[6] “WHITE PAPER: Information and Communications in Japan,”

4. 3 Discussion

[9] H. Yokota, et al., “Decentralized Micro-Mobility Management for

GWs do not need to have either a routing table or an ARP cache with the proposed architecture while GWs with a

Ministry

of

Public

Management,

Home

Affairs,

Posts

and

Telecommunications of Japan, 2003. [7] C. Williams “Localized Mobility Management Requirements,” draft-ietf-mobileip-lmm-requirements-03, IETF, March. 2003. [8] F. M. Chiussi, et al., “Mobility Management in Third-Generation All-IP Networks” IEEE Communications Magazine, Vol. 40, No. 9, pp.124-135, September 2002. Large Scale Mobile Networks,” WPMC, Oct. 2003.