Analysis of Load Dependency of Handover Strategies in Mobile Multiaccess Ambient Networks

in Mobile Multiaccess Ambient Networks

Petteri Poyhonen Nokia Siemens Networks Helsinki, Finland petteri.poyhonen@nsn.com Oliver Blume Alcatel-Lucent R&I Stuttgart, Germany

oliver.blume@alcatel-lucent.de

Abstract

Wireless operators incorporate multiradio access technologies aiming at expanding their customer base and benefiting from synergies of existing and planned infrastructure. The result is areas with overlapping radio access technologies, enabling users to select from a large pool of available connections, based on several criteria. In these increasingly complex scenar-ios the importance of handover decisions cannot be underestimated. We study three different handover decision strategies, focusing on the tradeoff between continuous connectivity, network utilization and per-formance, and the associated costs, and compare them to the optimal strategy. We motivate the need for a distributed algorithm, especially when considering the deployment of multimedia applications, and reflect on the effective network capacity as a function of the handover strategy employed and the permitted offline time for the active mobile nodes in the network.

satisfaction, end-to-end path optimality, and energy efficiency. Currently, the state of the art lets a mobile node (MN) to choose between different RATs based on rudimentary and static user and operator preferences. For example, 3G/UMTS MNs typically opt for 3G rather than EDGE or GPRS connectivity. Another static preference dictates that a wired LAN connection is preferred (when available) over a wireless one, which in turn is preferred over a 3G connection. Al-though such policies are commonly used, they can lead to suboptimal network use and unrealized potential. We attempt to shed some light on the performance of different handover (HO) decision strategies and their respective dependency on the load in a mobile multiac-cess ambient network environment. We employ simu-lation to measure the performance of different strate-gies under varying loads, and use linear programming to establish the corresponding upper bounds. The rest of this paper is organized as follows. Section 2 briefly surveys related work on multiaccess networks. Sections 3 and 4 describe our system model and HO decision strategies and present our results, respectively. Finally, Section 5 draws conclusions.

Daniel Hollos

Technische Universitat Berlin

Berlin, Germany hollos@tkn.tu-berlin.de

Ramon Aguero University of Cantabria Santander, Spain ramon@tlmat.unican.es

Haitao Tang

Nokia Siemens Networks Helsinki, Finland haitao.tang@nsn.com Kostas Pentikousis

VTT Oulu, Finland

kostas.pentikousis@vtt.fi

1. Introduction

The proliferation of radio access technologies (RAT) accompanied by advances in miniaturization and industrial design allowed vendors to market de-vices with several network interfaces. For example, quad-band phones with personal area network (PAN) interfaces are common, and will soon be equipped with wireless LAN (WLAN) interfaces. In a multiaccess environment it is an open research topic how to decide which interface(s) to use and which network to attach to when considering radio resource sharing, continuous connectivity, load balancing, security and AAA, user

2. Multiaccess Networks

The relevance of multiaccess networks has been -increasing recently. Though there may be several rea-sons for this, the most important consequence is that techniques, protocols and procedures currently used to manage access selection in wireless networks are deemed insufficient to cope with the challenges that arise. The research community is already addressing some of the challenges that emerge. Standardization

bodies are looking at this type of scenarios, e.g. IEEE 802.21 [1] is defining a Media Independent Handover service which allows information gathering from het-erogeneous wireless technologies, treating them evenly, and embracing multiple technologies and operators. The Ambient Networks (AN) project [2][3] has been developing a system architecture that considers multiple RATs during access decisions and can apply the appropriate handover protocol allowing for session continuity. In order to manage the system complexity, control functions of AN are split into dedicated func-tional entities (FE). Among these are the Multi-Radio Resource Management FE (MRRM) [9] and a mobility tool box comprising protocols, such as, Mobile IP [4] and Host Identity Protocol (HIP) [5]. MRRM manages different RATs and coordinates radio admission con-trol, radio link setup and signal strength measurements. In this work split several challenges arise, first of all that the FEs involved in decision making can be dis-tributed between several nodes. This implies that con-straint availability is restricted: For example, certain network-side constraints are not available at the MN (due to security, performance, and other limitations) and vice versa. Moreover, most FEs should not have to deal with intricate data from other FEs, and may oper-ate at different timescales.

Besides the AN concept, there are several papers addressing this sort of network deployments in differ-ent ways. Some of them focus in particular scenarios, e.g. analyzing the integration of 3G and WLAN tech-nologies [6], studying how different mobility solutions work on top of them [7], [8]. In [9], the benefits of us-ing different metrics to determine the most suitable access to improve the performance of user traffic are analyzed; in [10] the user-perceived performance bene-fits are analyzed, although tight cooperation and cou-pling of the involved access technologies is assumed, which may not be the case, e.g. when considering multi-operator scenarios. In addition, [9] and [10] pay attention to the access selection process, but do not consider HOs caused by end-user movements.

In contrast, this paper considers mobile users and examines the problem of contradicting objectives be-tween users and network operators. It describes how to combine such objectives and constraints originating from both the user and the network side. Typically, operators opt to distribute the network load among their base stations (BS) in order to avoid overload in some cells, as a primary objective; this, however, may result in an increased number of HOs, which can have adverse effects on user applications. In our model, the user’s primary objective is to select a suitable access but keep the number of HOs performed to a bare mini-

mum. We study the tradeoff between the number of users which can be supported and the number of HOs per MN. The model assumes the existence of a distrib-uted multiaccess network architecture, such as the one described in [11].

3. Handovers in Multiaccess ANs

In AN, HOs in multiaccess networks are realized via specific decision algorithms which will likely consist of decision points distributed among different entities of the network, all participating in the joint decision mak-ing process. Effectively, we face the problem of creat-ing a decision framework where (i) the information has to be kept distributed among the deciding participants (FEs); (ii) information availability depends on the loca-tion of FEs and, thus, no FE can maintain a complete view at all times; and (iii) information from different FEs changes at different time scales.

In this approach, MRRM is responsible for solving this problem by continuously monitoring for available accesses and generating a list called the detected (ac-cess) set (DS). This list is sent to other FEs, which in turn assign abstracted weights to each access included in DS reflecting their own priorities. MRRM collects the weights from all involved FEs and makes the final decision. This procedure avoids disclosing internal parameters which might be subject to confidentiality limitations. Note that MRRM, being the decision-making FE, can be operating in the MN and/or in the network side. Meeting constraints like policies, end-to-end QoS, and the required mobility protocols, MRRM stepwise reduces the DS into subsets called validated set (VS) and candidate set (CS) (explained next). Due to limitations in information sharing, different sets may be generated at the network and the terminal sides. Fi-nally one or more accesses are entered in the active set (AS) which is used in the communication session. Complete changes of the AS result in a HO.

3.1 Decision Strategies

We model three HO strategies which differ in the order of evaluating and filtering the information of the available accesses discovered (DS); the final selection is done at different places. In the terminal centric strat-egy, illustrated in Figure 1(a), the MRRM at the termi-nal builds the DS, and then creates its CS (CST) by assigning weights to the records using its own prefer-ences; then it sends this to the network which con-structs the corresponding CSN based on both the re-ceived CST and its own constraints. Note that cells with insufficient capacity are removed from this list

before applying any weights. The terminal uses the intersection of CST and CSN, and by means of a weighted sum, it selects the best accesses, thus creating a new AS. This may cause the MN to perform a HO.

Figure 1. HO strategies: simplified message exchange sequences.

In the network centric strategy, depicted in Figure 1(b), the terminal builds the DS but does not filter it before sending it to the network. The network creates the CSN (again the accesses which are overloaded are silently discarded) and passes it to the terminal which, applying its own constraints generates the CST. The final decision is taken at the network side, again, using the accesses that belong to both CST and CSN by ap-plying a (possibly different) weighted sum function, and sends it over to the MN for HO execution.

The former two strategies are compared to a legacy strategy that only applies signal strength to rank the connection requests. It checks whether there is enough available capacity to handle new access requests and rejects new ones if the cell is full. This strategy does not use inter-RAT policies for weighting, i.e. the policy of maintaining the currently active RAT and the net-work policy of load balancing and is the simplest from a migration perspective as it does not require distrib-uted decisions nor a network MRRM. It is expected to create less HO events, but will not yield the most net-work capacity gain.

In our simulations, described in the following sec-tion, we measure load in terms of traffic units (TU). We use two types of traffic sources, requiring 1 and 1.5 TUs respectively. Very roughly, this maps to MNs generating VoIP traffic, using two different types of codecs, say, G.729 vs. iLBC. In this sense, we can model different service requirements. The network centric strategy also applies a further network policy that favors cells mapping to the requested traffic type.

For each constraint (see Table 1), the accesses are evaluated and a constraint specific ordered access list is created as depicted in Figure 2. After all constraints are applied, the resulting access lists are used to construct the integrated access set, according to the constraints’ weights.

Table 1. HO constraints and weights in our model.

Legacy Terminal Network

centric centric

Constraints Signal Signal strength; Signal strength; on CST strength same RAT; same RAT;

preferred preferred traffic type traffic type

Constraints Cell Cell overload; Cell overload; on CSN overload load balancing; load balancing;

available traffic available traffic

type type

Apply Terminal-weights at N/A Network-Network Terminal

Figure 2. Access evaluation with four constraints.

4. Performance Evaluation

In our evaluation we take N MNs (N=100...1000) moving in a 1km×1km area, which is served by three different RATs. All in all there are 30 cells of type WLANa (RAT-1), 4 cells of WWAN (RAT-2), and 60 cells of type WLANb (RAT-3) in the 1 km2 area, which is entirely covered by at least one of the RATs, and there exist subareas with extensive overlapping. The MNs move according to a mobility model akin to the random direction model with the following parameters: max speed = 10 m/s, random ±90° direction change probability, and no idle moments. All algorithms are executed once per 0.1s, for a total simulated time of 120s.

We used MATLAB to evaluate the above-mentioned strategies; the decision methods are simulated in a way

that for every 0.1s time slot the criteria for all terminals are re-evaluated. We assess the three methods using the same simulation settings and compare the results against the optimal allocation, calculated by a linear programming (LP) technique, called Mixed Integer Programming (MIP). MIP finds one of the optimal so-lutions if it exists, and reports unfeasibility otherwise; therefore it is commonly used in network planning to obtain (theoretical) upper bounds on performance [12]. The optimal approach is not practical due to calculation power limitation for real-time optimization and huge amount of HOs to realize the optimum

Solving the following MIP each of the 1200 0.1s-time slot returns the best HO decisions for all MNs in the area for the objective of maximizing the overall network load:

max. ∑R∀i,bi,b,rAi,b,r

(1)

,r

subj.to ∑A∀b,ri,b,r≤1

∀i∈MN (2)

∑R∀i,r

i,b,rAi,b,r≤Cb

∀b∈BS (3)

where Ai,b,r is the decision variable, i.e. the result of

the calculation which is to be assigned by the MIP solver.Ai,b,r:={0,1} is set to one if MNi is connected to

BSb using traffic type r. The input parameter matrix

Ri,b,r:={1,1.5} includes the load generated by MNi if

connected to BSb using traffic type r, in terms of TUs; note that in the input parameter list Ri,b,r only the feasible (i,b,r) combinations are listed, i.e., when MNi is within coverage of BSb and can connect using traffic type r. Finally, Cb is the capacity of BSb in TUs.

The objective function (1) aims at maximizing the overall load the MNs generate: since Ai,b,r (set to one for each active connection) is multiplied by the pa-rameter Ri,b,r (expressing the corresponding load in TUs for the (i,b,r) combination), the sum in (1) gives the overall network load. Constraint (2) ensures that each MN is connected to at most one access point at a time, while constraint (3) limits access points to their capacity. The above MIP model can be solved using several solvers; we used the CPLEX mathematical pro-gramming optimizer in this study. We will refer to the results of this solution per time slot as “optimal” in the remainder.

4.1 Results

Figure 3 shows the average MN connectivity as a percentage (%) of the total simulation time for the op-

timal (MIP solution) and for the three strategies

(MATLAB simulation results). As N increases, the network becomes increasingly overloaded and for N=1000, MNs are, on average, connected less than half of the total operation time (120s). This result shows the large effect of the HO decision strategy on connec-tivity. In our model, the network centric method always outperforms both the terminal centric and the legacy strategies. However, as the optimal curve shows, there is still room for approximately 10% improvement.

As N increases, the average online time difference between the network centric and the legacy remains similar, due to the rigid decisions of the legacy strat-egy. In contrast, the terminal centric method is closer to the legacy one for under-loaded networks, i.e., it per-forms rather badly, but gets towards the network cen-tric method for high loads. The reason of this effect is that in an under-loaded network most if not all MNs can be supported, while in the heavily overloaded case there is little, if any, room for improvement.

100OptimalLegacyTerminal 90Networkem 80it enilno fo 70 egatnecre 60P egarevA 50 40 30 100 200 300 400 500 600 700 800 900 1000Number of MNs simulatedFigure 3: Average Percentage of Online Time (%) un-

der the number of the mobile nodes in the network.

The higher connectivity results in the terminal and network centric strategies come, however, at a high price, as shown in Figure 4 which illustrates the aver-age numbers of HOs per MN for varying N. For N=300, the network centric strategy causes almost three times more HOs per MN when compared to the terminal centric and the legacy strategies. This is a re-sult of our rudimentary load balancing scheme which –having the playroom in the under loaded network and the objective of balancing the cell loads– always moves the MNs back and forth between the overlapping cells based on the current cell loads. With increased load,

90TerminalNetwork 80Legacy 70NM re 60p sOH 50fo rebmu 40n egare 30vA 20 10 0 300 400 500 600 700 800 900 1000Number of MNs simulatedFigure 4. Average number of HOs per terminal during

the simulated time (120s).

this scheme has less available capacity to play with; therefore this side-effect is suppressed. We plan to ad-dress this limitation shortly.

Figure 5 presents the number of MNs that can be supported by the system vs. the average allowed dis-connection time per MN. It shows the effective network capacity as a function of the HO strategies and the al-lowed offline time percentage for the active MNs in the network. On average, the smaller the allowed offline time proportion, the larger the improvement the net-work centric method has over the terminal centric one. The maximum improvement is for the 0-10% range of offline time. Since the usual value of the required al-lowed offline time percentage for the MNs is smaller (or even much smaller) than 5%, this is a very impor-tant finding: an operator could increase its network capacity by more than 100% over the legacy case if the the appropriate HO decision strategy is employed. This could mean significant gains in business income for network service providers. The lack of rich network-side information in its decision strategy, the terminal centric strategy performs much worse than the network centric strategy, while it still provides roughly 20% more effective network capacity than the legacy strat-egy.

In all cases, the network centric strategy is the best among the three studied HO strategies. In addition, it is significantly better when requiring the allowed offline time percentage for the active MNs in the network to be smaller than 5% (as usually required in practice). The optimal curve shows that the room left for im-provement over the network centric method depends on the network load; however, in the accepted range (0-10%) it is around 33%.

Figure 6 shows the distribution of the traffic among the different traffic classes vs. the number of MNs

900NetworkTerminalLegacy 800Optimal 700dev 600res sNM 500fo rebmu 400N 300 200 100 0 10 20 30 40 50Allowed offline time for the active MNs [%]Figure 5. Average network capacity realized by the

three HO strategies under different offline time per-centage allowed for the active MNs.

800

700

ask_1.5 600

500

sNM fo re 400

bask_1.5_disconnmuN 300

ask_1.5_get_connask_1 200

100

ask_1_get_connask_1_disconn 0 100

200 300 400

500 600 700 800 900 1000

Number of MNs simulated

Figure 6. Requested- and admitted load (in TUs) to

MNs for the optimal strategy.

simulated, for the optimal HO decision strategy. We use the following labeling format for the curves: ask_X is the number of MNs requesting X TUs; ask_X_get_conn is the number of MNs requesting X TU and receiving (at least) X TUs; finally, ask_X_disconn is the number of MNs that requested X TUs but ended up being disconnected.

As one might expect (Figure 6), the two pairs of curves ask_1.5 & ask_1.5_get_conn, and ask_1 & ask_1_get_conn show that as N increases the admitted capacity decreases. The reason is that getting over-loaded, the optimal strategy drives towards admitting the lower traffic class, since then the probability is higher that it can be admitted and fill the ‘last’ free resources of some cells.

4.2 Discussion

Our traffic model involving MNs moving in a het-erogeneous radio cell topology shows that system us-

age and thus potential operator revenue depend on the accepted probability of offline time. It yields that the maximum simultaneously served traffic (the sum of the _get_conn curves in Figure 6) is limited even with opti-mum access decision strategy to 75% of the total cell capacity ∑Cb which is 570 in our scenario). With the studied realistic access selection algorithms only 30-50% of the total capacity can be used, when the accept-able offline time is below 5%.

Load balancing coupled by considering the traffic type parameter promise to be effective methods to in-crease the amount of traffic transported without sur-mounting the acceptable offline time and without hav-ing to install more cells. This requires that multiple entities in different locations participate in a distributed access selection. Generic interfaces and protocols for exchanging CSs and access selection constraints are currently an important item in the Ambient Networks system specification agenda.

5. Conclusion and Future Work

This paper studied three different handover decision strategies through numerical analysis and simulation. We focused on the tradeoff between network perform-ance and the costs, and compared them to the optimal strategy. The analysis shows that the effective network capacity depends on the handover strategy employed and the allowed offline time for the active mobile nodes in the network. The results indicate that, for a single operator scenario, a network centric handover strategy may achieve a significantly higher network capacity than the other handover strategies.

Of course, although our work seems very promising it has several limitations which we are actively ad-dressing. Next steps as we improve the simulation model will be to include the analysis of a broader set of scenarios, introducing different traffic models, mobility patterns and cell topologies, as well as different simu-lation frameworks will be used so as to check the cor-rectness of the results, as well as to extend them. Fur-ther, we are planning study scenarios with several competing operators where the network centric strategy cannot be executed by a central entity, and develop algorithms balancing better between the connectivity and the number of HOs performed.

Acknowledgment

This work has been carried out in the framework of the Ambient Networks project (IST 027662), which is partially funded by the Commission of the European

Union. The views expressed are solely those of the authors and do not necessarily represent the views of their employers, the Ambient Networks project, or the Commission of the European Union. The comments and ideas from people involved in the project's mobil-ity research are gratefully acknowledged. In particular we thank Ove Strandberg for comments & suggestions.

References

[1] IEEE 802.21, “Draft IEEE Standard for Local and Met-ropolitan Area Networks: Media Independent Handover Services,” March 2006.

[2] N. Niebert, A. Schieder, H. Abramowicz, G. Malmgren,

J. Sachs, U. Horn, C. Prehofer, and H. Karl. „Ambient Networks – An Architecture for Communication Net-works Beyond 3G,” IEEE Wireless Communications, 11(2):14–21, Apr. 2004.

[3] Ambient Networks Project: www.ambient-networks.org. [4] C. E. Perkins (Ed.), “IP Mobility Support for IPv4,”

IETF Request for Comments: 3344, August 2002.

[5] R. Moskowitz and P. Nikander, “Host Identity Protocol

(HIP) architecture,” IETF Request for Comments: 4423, May 2006.

[6] Apostolis K. Salkintzis, “Interworking Techniques and

Architectures for WLAN/3G Integration Toward 4G Mobile Data Networks,” IEEE Wireless Communica-tions, pp 50–61. June 2004.

[7] L. Ma, F. Yu, V. C. M. Leung, and T. Randhawa, “A

New Method to Support UMTS/WLAN Vertical Hand-over Using SCTP,” IEEE Wireless Communications, pp. 44–51, August 2004.

[8] W. Wu, N. Banerjee, K. Basu, and S. K. Das, “SIP-Based Vertical Handoff between WWANs and WLANs,” IEEE Wireless Communications, pp. 66–72, June 2005.

[9] P. Magnusson, F. Berggren, I. Karla, R. Litjens, F.

Meago, H. Haitao, R. Veronesi, “Multi-radio resource management for communication networks beyond 3G,” VTC Fall 2005 - IEEE 62nd Semiannual Vehicular Technology Conference, vol. 3, pp. 1653–1657, Sep-tember 2005.

[10] F. Berggren, R. Litjens, “Performance analysis of access

selection and transmit diversity in multi-access net-works,” Proc. 12th annual international conference on Mobile computing and networking, pp. 251–261, 2006. [11] M. Prytz and J. Gebert (Eds.), “Multi-Access & ARI

Design and Initial Specification,” Ambient Networks, Deliverable D2, December 2006.

[12] I. Kadayif, M. Kandemir, N. Vijaykrishnan, and M. J.

Irwin, “An integer linear programming-based tool for wireless sensor networks,” J. Parallel Distrib. Comput., vol. 65, no. 3, pp. 247–260, 2005.

[13] J. Wong, M. Neve, and K. Sowerby, “Optimisation strat-egy for wireless communications system planning using linear programming,” IEE Electronic Letters, vol. 37, pp. 1086–1087, 2001.