A Hybrid Load Balance Mechanism for Distributed Home Agents in Mobile IPv6

A Hybrid Load Balance Mechanism for Distributed Home Agents in Mobile IPv6 1 Hui Deng 2Xiaolong Huang 3Kai Zhang 3 Zhisheng Niu 1Masahiro Ojima 1R&D Center Hitachi (China) Ltd. Beijing 100004, China 2Dept. of Electrical Engineering, UCLA, Los Angeles, USA 3Dept. of Electronic Engineering, Tsinghua University, Beijing 100084, China Abstract- Mobile IPv6 is a key technology in IPv6 to support the mobility of wireless communication terminals. In Mobile IPv6, Home Agents (HAs) are responsible for the registration of Mobile Nodes (MNs) in the home network, and tunneling the data packets to the MNs when the MNs are not reachable through its home IP addresses. However recent research shows that the traffic bottleneck could be formed at a HA. When the HA experiences high intensity of the tunneled traffic and the MH registration information. In this paper, we propose a hybrid load balance mechanism which takes account of not only the MN registration information but also the tunneled data traffic information to effectively release and prevent the formation of the traffic bottleneck at the HA. We show that the proposed mechanism can be implemented in Mobile IPv6 without changing the protocols of the communication between HAs and MNs in IETF MIPv6 draft. Simulation shows that our proposed algorithm can reduce the traffic delay substantially and the buffer requirements during the tunneling traffic phase in Mobile IPv6. I. INTRODUCTION Recently, great progress has been made in the wireless communication networks and the Internet. Mobile IPv6 [1] is considered to be one of the key technologies for realizing the integration of the wireless communication and the Internet, which enables the seamless communication between wired and wireless networks. In Mobile IPv6, a HA should maintain the registration information of the MN, which is away from its home network. The HA is also responsible for receiving the IP datagram on behalf of the currently registered MN and tunnel datagram for delivery to the MN. When the number of serviced MNs rises significantly, the packets will queue up at a HA, causing long delays of the data traffic and the registration process. Under certain traffic conditions, such as supporting multimedia applications for a large amount of MNs in Mobile IPv6 networks, a traffic bottleneck can be formed at the HA by the tunneled data traffic destined to the MN and the overhead of the MIPv6 routine tasks at the HA. The bottleneck can cause large delays of the data traffic and may force the HA offline. Several solutions [2-4] have been proposed to solve the traffic burden problem at HA by deploying multiple HAs in the home network, so that the traffic can be shared among the HAs. Jue [2, 3] designs and analyses a replicated server architecture in which multiple HAs are used to provide mobility support. Load balancing schemes are designed so that HAs are dynamically assigned and changed according to the traffic at each HA. They have studied the performance of three selection schemes: random, round-robin, and join the shortest queue (JSQ), and three transfer s policies: timer-, counter-, and threshold- based independently. However, the purpose of their study was to validate the numerical results obtained using their analytical model rather than focusing on the real case in Mobile IPv6, and therefore it has certain limitations and restrictions. Their results are also sensitive to apparently insignificant parameter changes. The solution also cannot prevent the traffic burden in advance. Vislache [4] presented the study of the performance of multiple HAs networks by conducting a series of simulations, and introduced a realistic double-threshold load balancing policy. However this work also does not take account in IETF Mobile IPv6 draft case, where the tunneled traffic only occurs in a short period of time although it may frequently occurs. In Mobile IPv6, the dynamic home agent address discovery (DHAAD) provides a practical base to solve this problem. In Mobile IPv6, DHAAD gives the MN flexibility to select a home agent besides manually assigning a HA to a M N. Using DHAAD, M Ns can retrieve the addresses of the HAs on its home link by sending ICMP Request packets to its home network. When the HA in the home network receives the ICMP requests, the HA gives the MN a list of the HA IP addresses by sending ICMP Reply packets to the MNs. The HA list is ordered by the HA preference. The MN will select the HA with the highest preference in the HA list to be its home agent. DHAAD illustrates to deploy multiple HAs and to dynamically change the HA for a MN is applicable through ICMP messages. However in DHAAD, a MN only sends ICMP request packets in rare situation, such as, HA failure, MN reconfiguration. The HA list cannot be sent to a MN frequently for HA reassignment, since the overhead of the ICMP packet with a HA list inside could be significant. All of the above solutions ignored preventing the formation of the traffic burden in advance and did not give the solution to realize the Load Balance mechanism based on the IETF Mobile IP standard (that is to say, those solution are analytical model and away from realistic), and Mobile IPv6

are seldom considered compared with Mobile IPv4. As we will show in this paper, our proposed Hybrid Load Balance mechanism with the deployment of multiple Home Agents in a home network will work on the tunnel traffic information and the registration information at the each HA to release the traffic burden and also prevent the traffic burden. The proposed mechanism can be embedded in the Mobile IPv6 standard using DHAAD. So it can realize rather than analytical model and it can prevent from traffic burden in advance. The entry of the traffic load table is: 1. Home Agent Address The HA address is the IP address of the base station where the HA resides. 2. Queue Size (Fig. 2) The traffic load indicates the buffer size at a HA. When the buffer size of a HA is lower than a threshold, the buffer size is considered to be LIGHT. This paper has following sections. Section II gives the proposed Hybrid Load Balance mechanism. Section III explains the simulation in NS2 and gives the simulation result. And finally Section? gives the conclusion of this paper. II. HYBRID LOAD BALANCE MECHANISM The following Figure 1 gives the topology layout to deploy the Hybrid Load Balance mechanism for distributed home agents Figure 2. Traffic Load Table 3. Registered MN number at each HA. Figure 1. Triangle routing traffic burden A. Initialization of the home network In the proposed Hybrid Load Balance mechanism, the home network is composed of multiple Mobile IPv6 HAs and multiple MNs. Each HA in the home network is attached with an access router. When the MNs reside in the home network, the HAs do not execute any HA tasks. The HA assignment of the MNs in the home network can be either evenly assigned among the multiple HAs or unevenly assigned. Whether the HA assignment is even or not would neither arbitrarily affect the original traffic burden problem nor affect the performance of the proposed Hybrid Load Balance mechanism. B. Traffic load table in the home network The proposed Hybrid Load Balance mechanism shares the traffic information among the HAs in the home network to make decisions of Home Agent Reassignment. To acquire and maintain the traffic information, each HA maintains a so-called Traffic Load Table. The Traffic Load Table has records to indicate the traffic load level of all HAs in the home network. The HA should monitor its queue size and the registered MN number. Each HA periodically broadcasts its traffic load advertisement to all the other HAs in the home network. The traffic load advertisement has the same fields as in the traffic load table. Upon receiving the traffic load advertisement from other HA, the HA should record the traffic load advertisement into the Traffic Load Table. The HA keeps the Traffic Load Table sorted in a non-ascending order of the traffic load field, unless the traffic load is LIGHT. For the LIGHT HA, the traffic load table is sorted in a non-ascending order of the registered MN number. In our proposed Hybrid Load Balance mechanism, the Queue Size field is used to make decisions for HA reassignment to release the traffic burden, while the registered MN number field is used to prevent the formation of the traffic burden. C. Home Agent Reassignment for the registered mobile hosts

Figure 3. Home Agent reassignments In the proposed Hybrid Load Balance mechanism (Fig. 3), at each HA, a timer is attached for each entry in the binding update cache. When the timer is time out, the MN corresponding to the entry is considered to be eligible for HA reassignment. The HA may select a new HA in the Traffic Load Table for the timeout MN according to our HA reassignment algorithm. If a new HA is assigned to the timeout MN, the HA actively sends out an ICMP Reply message to the MN without the reception of any ICMP Request message. Different from the standard ICMP reply packet, the ICMP here should only contain one HA, which is the newly selected HA, other than contains a list HA. By receiving this ICMP message, the timeout MN should compare the indicated HA with its old HA. If the indicated HA in the ICMP Reply message is different from the old HA, the MN should modify its HA field and register at the new HA by sending a binding update message to the new HA IP address. By using the ICMP messages in the DHAAD mechanism, our proposed Hybrid Load Balance mechanism can be implemented in the IETF Mobile IPv6 draft without any changing of the protocols of the communication between HAs and MNs. The frequency to select a new HA for the MN is a tradeoff between the HA handoff frequency and the load balance performance. The HA should not frequently select a new HA for the registered MN, because the HA handoff induces extra control traffic and delays the traffic forwarding to the mobile nodes. Thus only a very busy HA or a potentially very busy HA should proceed to the HA handoff. Randomly select a HA with LIGHT Queue. ELSE IF (Self Queue Size is top 10% in the traffic table) THEN Randomly select a bottom 10% home agent in the traffic table ELSE THEN IF (my registered MN number is top 10% in the traffic table) THEN Randomly select a bottom 10% home agent in the traffic table. In the HA Reassignment algorithm, only one of the most busy HAs can select a new HA for its registered MH. Thus the new HA assignment does not take place frequently. In IETF Mobile IPv6, a MN only needs the HA to tunnel the data traffic before the Correspondent Binding Update Procedure take place, when the MN is moving form one network to another network. Thus a HA who has more number of registered MNs is more likely to experience tunnel traffic because more M Ns potentially will move from one network to another network. The proposed Hybrid Load Balance mechanism can force a HA to start the new HA assignment even though the HA does not experience much traffic, so that the future traffic burden could be prevented statistically. III. SIMULATION We implemented the Hybrid Load Ba lance mechanism and the DHAAD of Mobile IPv6 in Network Simulator 2 (NS2). We also partially implemented the router advertisement method of Mobile IPv6 in NS2; so that the Mobile IPv6 tunneled traffic has no problem going through the LAN. A process queue is added to each home agent to emulate the processing overhead of the Mobile IPv6 tasks. The Mobile IPv6 protocol in NS2 is re-written based on MobiWan [6]. When selecting a new HA, the new HA should be one of the most released HAs in the Traffic Load table. There are two fields in the traffic load table should be considered in the home agent selection algorithm. One is the Queue Size field, which indicates the current traffic load. Another one is the registered MN number, which indicates the potential traffic load in the future. The HA should prevent from having too many registered MNs, so that the future traffic burden formed by the tunneled traffic for the registered MNs could be prevented. The new HA Reassignment algorithm is as follows. Algorithm. HA Reassignment IF (Self Queue Size > LIGHT) THEN IF (Other HA Queue Size < LIGHT) THEN Figure 4. Basic Simulation Configuration The basic network configuration is shown in Figure.4. The Correspondent Node (CN) sends the data traffic through the UDP packets to the MN. When the MN resides in the home

network, the traffic goes through the gateway directly sent to the MN. When the MN roams out of the home network, the data packets will be sent to it HA, and will be tunneled to the foreign network where the MN resides. In the first scenario, three MNs move out side of the home network consecutively, when the CN sends a Poisson arrival traffic session to all three MNs. The aggregated traffic arrival rate is set as same as the processing rate at each HA. At the beginning of the simulation, HA1 is manually assigned to all three MNs. In the second scenario, we concentrate on investigating the parameters in the Hybrid Load Balance mechanism. The main parameter of the proposed mechanism is the ratio of the Traffic Table update rate over the traffic arrival rate. In the second scenario, there are 20 MNs in the home network. 10 of them will move from one network to another network simultaneously. 5 MNs are assigned to each HAs. The aggregated traffic intensity of 5 MNs are slightly less than the processing rate of a HA. Thus the traffic situation is close to saturation. We adjust the Traffic Table update rate to see the overflow rate of the buffer at each HA with or without the Hybrid Load Balance mechanism. The result is shown in Figure 7. Figure 5. Buffer size performance Figure 5 shows the queue size of the process queue at each HA with or without the Hybrid Load Balance mechanism. The result shows when the HA experiences a saturate traffic situation, the proposed mechanism can share the traffic of the saturated HA, while the queue size at each HA is largely reduced. Figure 6 shows the delay experienced by each traffic session with or without the Hybrid Load Balance mechanism. The big oscillation of the delays at the end of the triangle routing is because the delay difference of the tunneled data traffic and the data traffic sent directly to the new Care of Address of the MN, after the Correspondent Binding Update procedure between the CN and the MNs. The simulation result shows the traffic delay is largely reduced by the traffic sharing mechanism during the tunneling traffic phase. Figure 7. Over flow rate of the buffer size during the tunneling traffic phase The result shows when the traffic advertisements rate is slightly slow than the traffic inter-arrival rate, the HA experiences least of traffic overflow. It is because the Traffic Table update rate can sufficiently provides the current traffic information at each home agent. When the Traffic Load Table update rate is slow, the Traffic Table cannot reflect the real state of the traffic distribution among the HA, thus the HA Reassignment cannot select a suitable HA for the timeout MNs. When the Traffic Load Table update rate is faster than the traffic inter-arrival rate at each HA, it may induce too much traffic so that the traffic within the LAN will affect the HA performance. Figure 6. Traffic delay of each session

Hybrid Traffic Load mechanism is based on the existing DHAAD protocol in the existing Mobile IPv6 standard. The mechanism can be included in the Mobile IPv6 standard without changing the protocols of the communication between the HAs and MNs. The simulation in NS2 shows that our proposed algorithm can reduce the traffic delay substantially and the buffer require ments during the tunneling traffic phase in Mobile IPv6. ACKNOWLEDGEMENT The authors would like to thank Hitachi Ltd. for the support in this work. Figure 8. Mean buffer size of 4 home agents during the tunneling traffic phase Figure 8 shows the mean buffer size of the 4 HAs in the second scenario. It is shown that the buffer size has been largely reduced. Figure 9 shows the average value of the buffer size variance over the buffer size mean during the tunneling traffic phase. The variance of the buffer size can indicate the stability of the Hybrid Load Balance mechanism in a statistical view. REFERENCES [1] D. B. Johnson, C. Perkins, and J. Arkko, Mobility support in IPv6 <Draft-ietf-mobileip-ipv6-21> IETF, 2002. [2] J. Jue and D. Ghosal, Design and Analysis of Replicated Server Architecture for Supporting IP-Host Mobility. ACM Mobile Computing and Communications Revue, 2(3):16-23, 1998. [3] J. Jue and D. Ghosal, Design and Analysis of Replicated Server to Support IP-Host Mobility in Enterprise Networks, IEEE International Conference on Communications v3:1256-1260, 1997. [4] A. Vasilache, J. Li and H. Kameda, Load Balancing Policies for Multiple Home Agents Mobile IP Networks, IEEE International Conference on Web Information Systems Engineering v2:178-185, 2001. [5] N. Montavont and T. Noei, Handoff Management for Mobile Nodes in IPv6 Networks, IEEE Communication Magazine, 40(8):38-43, 2002. [6] T. Ernst, MobiWan: NS-2 extensions to study mobility in Wide-Area IPv6 Networks, available through http://www.inrialpes.fr/planete/pub/mobiwan/ Figure 9. Variance/Mean of the buffer size of 4 home agents during the tunneling traffic phase IV. CONCLUSIONS In this paper, we proposed a Hybrid Load Balance mechanism deploys multiple HAs to share the traffic in the home network. The proposed mechanism takes in account of not only the traffic information at each HA but also the registration information at each HA to release and prevent the traffic burden. HAs are dynamically assigned to the MNs when necessary. In particular, a HA will be reallocated when there is a considerable traffic load difference between HAs, thus the HA assignment frequency is reduced. The HA reassignment also processes the registration MN information to probably prevent the potential traffic burden. Finally the