Dynamic resource management for QoS ... - Semantic Scholar

Computer Communications 29 (2006) 730–740 www.elsevier.com/locate/comcom

Dynamic resource management for QoS provisioning over next-generation IP-based wireless networks Sotiris Maniatis*, Eugenia Nikolouzou, Iakovos S. Venieris School of Electrical and Computer Engineering, National Technical University of Athens, 9 Iroon Polytechniou Street, 157 73 Athens, Greece

Abstract A significant issue in current research pursuits is the formulation of the requirements and basic design options for the next-generation wireless network architecture. The next-generation of wireless systems will support a diverse set of access technologies and mobile devices, formulating a broad heterogeneous environment with increased requirements on network support operations. It is expected that the demanding breed of multimedia applications will even more considerably require Quality of Service support throughout the end-to-end path. This paper first provides a tutorial approach on next-generation wireless network architectures and more specifically on end-to-end QoS provision. We claim that dynamic resource management in the Core Network is a necessity due to the increased heterogeneity of the new environment. We subsequently present our proposal regarding a dynamic resource management scheme that is based on the concept of the Resource Pools. The Resource Pool concept is deeply analysed within the paper and simulation results prove its correctness and appropriateness. q 2005 Elsevier B.V. All rights reserved. Keywords: Next-generation wireless networks; Quality of service; Dynamic resource management; Network services

1. Introduction The next-generation wireless networks will be principally formed through the evolution and convergence of current mobile communication systems and the IP technology. The foreseen architecture will retain the well-known bi-level structure, consisting of a multi-domain Core Network (CN) offering IP connectivity and services, and a set of wired and wireless Access Networks (AN) offering the last mile connectivity services to mobile users. The convergence concept has gained significant attention during the last years in the research, industrial and standardisation community, leading to a number of convergence scenarios and propositions, like in Refs. [1–3]. The fundamental characteristic of next-generation networks is heterogeneity. Heterogeneity is the result of the different operations and capabilities of the multiple converged access networks, the reconfiguration and adaptation capabilities of end-user * Corresponding author. Tel.: C30 210 772 2318; fax: C30 210 772 1092. E-mail addresses: [email protected] (S. Maniatis), enik@telecom. ntua.gr (E. Nikolouzou), [email protected] (I.S. Venieris).

0140-3664/$ - see front matter q 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2005.07.016

devices and applications, as well as the interworking model resulting from the various handover scenarios and the existence of multiple providers. The architecture of next-generation wireless networks will therefore have to deal with heterogeneity. The latter imposes strict requirements that have to be satisfied in order to provide a smooth and seamless service. These requirements affect the basic operations of the core network, including mobility management, network resource and Quality of Server (QoS) management, and overall AAA operation control, among others. This paper focuses on network resource management for QoS provisioning in a heterogeneous next-generation wireless network environment. The main motivation of our work is that the heterogeneous nature of such environment necessitates the existence of a dynamic resource management layer in the core network of the architecture, as briefly explained hereafter. In third generation (3G) or previous systems, the CN is not considered the bottleneck of the overall network. In such networks the bottleneck is always the pure wireless part of the AN, allowing the operators to properly dimension the remainder network (the wired part of the AN as well as the CN) so that minimum congestion will occur in that. Dimensioning here does not necessarily mean

S. Maniatis et al. / Computer Communications 29 (2006) 730–740

over-provisioning as one could claim, but instead a more or less static management of CN resources, according to predicted traffic patterns, and based on the standard CN’s QoS capabilities. This is eventually true for any homogeneous wireless network that gives to the administrator the opportunity to study the traffic patterns, predict the traffic demands and at last appropriately dimension the network. Dimensioning in homogeneous wireless networks, although possible, is indeed very demanding and difficult, mainly due to user mobility. Dimensioning is, however, far more difficult for heterogeneous networks that expose dissimilar traffic patterns stemming from the multiple available access networks. Moreover, user mobility does not only entail an intra-system handover, but it may involve an inter-system one, resulting in considerable adaptation of the traffic involved. Furthermore, with the advent of ad hoc wireless networks, which can be configured on demand, dimensioning and provisioning of the core network becomes even more intense. Due to the aforementioned facts, dynamic resource management in the core of next-generation wireless networks appears to be a necessity. There is common consensus on the fact that CN will be a pure IP-based network, meaning that all CN operations will be based on mechanisms and protocols developed and used in the IP world. The chief standardisation body regarding IP is the Internet Engineering Task Force (IETF) [4]. Concentrating on QoS, which is the focal point of this paper, IETF has basically defined two frameworks: the Integrated Services (IntServ) and the Differentiated Services (DiffServ). The former offers QoS guarantees with the aid of the RSVP protocol but exposes scalability problems, while the latter provides only soft-QoS-guarantees. The introduction of the Bandwidth Broker (BB) concept by the Internet2 QBone [5] was an effort to basically cater for resource management and admission control over DiffServ networks. However, the BB architecture has not been standardized in IETF. As it will be presented in Section 2, current research trends favour the BB-enhanced DiffServ framework compared to IntServ and RSVP. Our work, presented in this paper, is also based on this framework. The dynamic resource management presented in this paper is based on the concept of Resource Pools. Resource Pools (RPs) try to overcome the scalability problems of the centralised BB concept, by introducing a distributed and highly scalable resource management scheme, and cater for a dynamic distribution of resources to the different heterogeneous access networks connected to the CN based on real traffic demands. The fundamental idea is to organize a number of Edge Routers1 (ERs) sharing a common bottleneck element into groups, which are called RPs. Those groups will provide a dynamic and efficient way for sharing 1 In the context of this paper, Edge Router is considered the network element which interconnects an access network with the CN.

731

and shifting the available resources between RPs based on real traffic demands. In the context of the RPs model, some algorithms are proposed to cater for dynamic (re)distribution of resources among RPs, in case the real resource demand is not met by the initial provisioning scheme. Those algorithms derive from the AQUILA [6] framework, which comprises a wired network, but are further enhanced to encounter the requirements introduced by the heterogeneous wireless segments. The rest of the paper is structured as follows. Section 2 presents a tutorial-in-nature approach for the next-generation wireless network architecture, identifying the basic network elements and operations, discussing the control layer needs for QoS provisioning and the network services provisioning issue, and lastly summarizing the state of the art and current trends in end-to-end QoS. Section 3 explains the dynamic resource management mechanism, while Section 4 presents some simulation results that prove its appropriateness. Finally, Section 5 contains the conclusions and future work.

2. Next-generation wireless network architecture There is a common consensus about the fundamental building blocks of the next-generation wireless network architecture. A common IP-based Core Network provides the mobile devices with the basic IP connectivity and uses only native IP protocols (IETF-based) for every operation: network address assignment, network management, mobility, quality of service, and AAA, among others. The Core Network retains the functionality of the well-known 3G Core Network, but it can be considered an evolution of the latter, since the intention is to decouple the wireless access technologies from the core network serving them. The various access segments will connect to this unified core through a generic interface (similar to Iu in 3G terminology), communicating with special routers that will have the ability to perform the adaptation functions needed for the underlying access technologies. We have to stress here that the Core Network will be able to serve not only wireless access segments but also wired ones. Fig. 1 depicts the envisioned architecture of the next-generation wireless environment. The morphology of the Core Network follows the typical IP network setup. Special Edge Routers (ERs) interconnect the various access segments to the core. Apart from performing access-specific adaptation functions, these routers operate as typical edge routers in the IP world. Note here, that instead of employing specialised ERs, an interworking unit could be provided at the edge of each access segment to perform the adaptation functions. In this case, ERs are conventional edge IP routers. The Core Network additionally consists of Core Routers (CRs) that interconnect ERs and Border Routers that provide the connectivity to external IP networks. In this sense,

732


Fig. 1. Next-generation wireless network architecture.

the whole network depicted in Fig. 1—consisting of the multiple access segments and the evolutionary CN— comprises an administrative domain (authority) or autonomous system. In order to satisfy the QoS requirements of mobile users, the next-generation wireless network architecture must provide a QoS framework, similar in notion to the current 3G one. Although heterogeneity complicates the overall QoS procedure, it is certain that the following essential elements of any QoS framework must be provided: a control layer, a data layer, and a resource management layer. The control layer specifies the mechanisms and protocols that allow the QoS requirements for a particular application or session to be specified and transported along the end-to-end path. The latter may include more than one domain. The data layer specifies the mechanisms that actually materialise the QoS provisioning processes within each domain along the e2e path. The data layer mechanisms are usually based on the concept of the network service (or traffic class), which is further discussed in Section 2.1. The resource management layer is typically responsible for the efficient

allocation of resources among several links, in order to maximise utilisation and minimise packet loss or rejections. Each one of the three elements has received a lot of attention recently, as will be discussed in a subsequent subsection. In this paper, we focus only on resource management, and particularly on the dynamic resource management needed in the core network. Radio resource management (RRM) and common RRM (CRRM), performed within each wireless access segment is rather important in such environment. End-to-end QoS provisioning has always been a complex task, at least from the network perspective, since the e2e path usually involves different domains that normally employ different QoS frameworks that have to interoperate. The heterogeneity of the next-generation wireless environment definitely further complicates the e2e QoS provisioning process. On the other hand, the fact that the core network is purely IP-based may alleviate the inter-domain QoS provisioning at core network edges. Fig. 2 tries to depict and explain these issues. Looking at the figure from top to bottom, one can say that in order to establish the end-to-end QoS service between the Mobile Device (MD) and the Correspondent Node (CN), three basic steps must be executed: (a) MD and CN exchange application-level signalling in order to establish the session(s) needed for the application to begin. For example, a video-conference application may employ the SIP protocol to initiate the call. The protocol employed at the application level may exchange network-level QoS parameters, although this is not always desirable due to the requirement to retain a clear, independent and flexible separation among applications and the network. Nevertheless, not all application signalling protocols are capable of negotiating QoS parameters at the network level. On the other hand, this implies that a proper translation is needed from application level QoS requirements to network QoS semantics. (b) MD and CN exchange QoS signalling at network-level, in order to allocate the requested resources and employ

Fig. 2. -End-to-end QoS control and mapping requirements.


the appropriate QoS provisioning mechanisms along the end-to-end path. Although not depicted in the figure, QoS signalling may be exchanged by appropriately configured proxies instead of end hosts. Each domain along the e2e path must be able to parse the QoS requirements carried by the e2e protocol and map these requirements to domain-specific network-level QoS semantics, in order to perform admission control and employ its individual traffic control and resource management mechanisms to support the QoS requirements. Regarding domains that follow the 3G or nextgeneration wireless network architecture, this step actually means to map the network level QoS requirements into their individual Core Network QoS semantics, since the CN actually employs these mechanisms. Mapping to CN QoS semantics is normally easy, since CN technologies usually utilise IP-based QoS frameworks. For example, in UMTS, the standards propose a DiffServ-based architecture to implement the CN Bearer Service.2 Moreover, in nextgeneration wireless networks, the CN is actually a pure IP network, which further alleviates the mapping process. (c) Domains that follow the 3G or next-generation wireless network architecture must map the network level QoS requirements into their individual Access Network QoS semantics. This task is specific to each access technology and usually means setting up and configuring the appropriate radio channels. Since, next-generation wireless networks make use of more than one AN technology and may employ CRRM techniques, it is evident that this task is complex enough, particularly in comparison to 3G architectures. Summarising the aforementioned discussion, e2e QoS provisioning needs three mapping stages: (a) Application level QoS requirements to network level QoS semantics. This stage should produce a generic (and eventually standardised) network QoS specification, in order to be transferred by the e2e QoS protocol. Considering that IP will be the common denominator of all future networks, this generic specification should be based on semantics defined by the IP world. (b) Generic network QoS specification to domain-specific Core Network QoS semantics. The adoption of the next-generation wireless architecture, which employs a pure IP CN, facilitates this stage. Considering an IPdominant world, this task mainly requires only 2 The 3G QoS architecture is expressed in Bearer Services (BS), which include all aspects that enable the provision of a contracted QoS, like the control signalling, user plane transport and QoS management functionality, among others. Although the UMTS standards suggest the usage of the DiffServ framework for the provision of the CN BS, they do not actually propose a specific implementation.

733

fine-tuning among different sets of network services defined at each domain (see Section 2.1). (c) Domain-specific network-level QoS specification to Access Network semantics. This stage mainly requires configuring properly the Access Network to accommodate the QoS requirements. The heterogeneity of next-generation wireless networks, coupled with CRRM techniques and inter-access technology handoffs, significantly raise the difficulty of this stage.

2.1. Network service provisioning In order to provide end-to-end QoS, some kind of traffic control and resource management must exist in all administrative domains that constitute the end-to-end path. Each domain employs its own mechanisms and defines its own parameters for traffic control and resource management. It is common that a QoS model is based on the notion of a network service (or traffic class). According to Ref. [7], a network service is ‘the definition of semantics and parameters of a specific type of QoS’. Thus, the materialization of QoS provisioning within one domain is based on the classification of the traffic to a set of network services. The most notable instances of existing set of network services are those of IntServ [8,9], UMTS [10], and ITU-T [11]. Each network service accommodates applications with similar QoS requirements by defining a specific forwarding behaviour to its packets and the appropriate combination of bounds on performance metrics. To be more specific, IntServ comprises three network services: guaranteed, controlled-load and best-effort. Under the Universal Mobile Telecommunications System (UMTS), four network services are defined: Conversational, Streaming, Interactive, and Background, which try to reflect the needs of the corresponding major types of applications. The same observation can be made for the six services defined recently by ITU-T in recommendation Y.1541. On the other hand, the DiffServ paradigm does not delineate a specific set of services, but it only provides a rough outline of how services in DiffServ could be defined. From the discussion above, it is apparent that a significant issue in next-generation wireless network architectures is the definition of the proper set of network services to be supported. One way to approach this issue is to follow the 3G paradigm. In other words, one could adopt the set of network services defined in UMTS. However, the heterogeneity of next-generation networks necessitates the enhancement of this set of services according to a few complementary characteristics, apart from mere QoS traffic characteristics, like user mobility, and inter-access technology handoff probability, among others. The key idea is to offer different (better) support for those sessions that expose high mobility patterns. To be more specific, it would be desirable to differentiate among a VoIP session happening

734


on the road (wireless mobile), from a VoIP session taking place at an office environment (wireless static). The former requires a more intense support from the network, in terms of both traffic control and resource management. At this point, the notion of mobility prediction is a significant fact that must be taken into consideration in the implementation of QoS provisioning at next-generation wireless network environments. The aforementioned issues are discussed in Refs. [12,13]. In Ref. [12], the authors present the MoDiQ service model, which is based on the Differentiated service framework, and proposes separate services for mobile and non-mobile sessions in order to enhance the system’s resource utilization in environments where mobile and non-mobile users exist. For example instead of one premium service two different are provided, namely the Mobile Premium Service and the Portable Premium Service, where the first provide support for assurances on the handoff success probability, while the second does not. Moreover, Ref. [13] presents another approach where velocity and call duration concepts are introduced for QoS provisioning and call admission support, which exhibits enhanced channel allocation and low call dropping probability. 2.2. Current trends in end-to-end QoS in heterogeneous wireless environments In order to achieve the goal of providing high-quality services in next-generation wireless networks, it is necessary to implement new techniques that can guarantee Quality of Service when considering the limitations imposed both by the end-user and the network. There have been many research approaches [14–21] trying to provide end-to-end QoS guarantees over heterogeneous networks, as briefly discussed hereafter. In Ref. [14], the main QoS challenges for the seamless support of different categories of hyper handover are summarized. The paper presents a QoS framework that includes a three-plane network infrastructure and a terminal-based hierarchical policy management system. A network manager and bandwidth broker based QoS framework is proposed to describe how the framework supports the end-to-end QoS provision adaptively and seamlessly. In Ref. [15], a QoS architecture in a heterogeneous radio access network is presented. Authors argue that policy-based service negotiation is an important component of the bandwidth broker and of the control plane, in order to provide a flexible endto-end QoS management in B3G systems. Furthermore, they identify the following significant issues: the need for a policy decision point at the radio bearer service entity, along with the interaction of service negotiation and mobility management protocols. The Resource Brokerage scheme is discussed in Ref. [16] where the interfaces and modules of the RB process are presented. Authors in Ref. [17] present a novel end-to-end QoS architecture that enables seamless services over heterogeneous wireless access networks. They

discuss the main architectural approaches and design issues of mobility-aware QoS signalling in IP networks. Then they introduce a QoS signalling architecture that integrates resource management with mobility management. In Ref. [18], a DiffServ resource allocation architecture is proposed for the evolving wireless mobile Internet, identifying two major challenges in establishing a wireless mobile Internet: support of fast handoff and provision of quality of service (QoS) over IP-based wireless access networks. The proposed registration-domain-based scheme supports fast handoff by significantly reducing mobility management signalling. The registration domain is integrated with the DiffServ mechanism and provisions QoS guarantees for each service class by domain-based admission control. Authors in Ref. [19] investigate the design of a wireless network architecture that exploits user profiles, such as the location, the velocity (both speed and direction), and the resource requirements of the mobile device, to maximize network efficiency and provide better QoS to different classes of users. The key underlying primitive of the architecture is the use of both real-time and aggregate user profiles to perform advance resource reservation in the handoff target cells of the wireless cellular network. In Ref. [20], the authors consider the use of IP-level QoS signalling as a key component to support the end-to-end QoS for various applications in the nextgeneration of mobile systems. They propose a small set of application programmer- and wireless-link-friendly IP QoS parameters and illustrate their use in a specific handover situation. In Ref. [21], the authors propose a hierarchical architecture for both mobility and QoS support in IP-based wireless networks. The proposed architecture has several advantages and provides excellent solutions to the problems raised by mobility and by the wireless environment. In QoS provisioning, they enable the end-to-end QoS guarantee by using the resource reservation protocol (RSVP) signalling. In particular, the RSVP aggregation technique is used to avoid the scalability problem. It is also worth mentioning the work conducted under the Next Steps in Signalling (NSIS) [22] IETF charter. The NSIS framework investigates the requirements, architecture, and protocols for QoS signalling across different network environments. The main idea of NSIS is to have a layered model, where the lower layer, called the signalling transport layer (NTLP), offers generic signalling functionalities, while the upper one, the signalling application layer (NSLP), is application-specific. This subdivision caters for the necessity to provide a general model, applicable in any part of the end-to-end path and valid to signal other than just QoS, even though NSIS is mainly focused on QoS. The main function of the NTLP is to forward signalling messages independent of the underlying network locating the appropriate next hop to address the signalling messages and identifying to which flow the messages belong. The NSLP uses the services offered by the NTLP to try to serve the main concern, which is QoS provision.


Another very important area of study deals with techniques for resource management in wireless networks driven by the aspiration to improve system capacity and performance under a specified set of constraints. In Ref. [23], an overview of radio resource management algorithms is presented, which include handoff, admission control, channel allocation and power control. Some algorithms combine two or more radio resource management tasks (mainly dynamic channel assignment and handoff) to provide better overall performance. An example is the admission control based on traffic prioritisation, between new calls and handoffs. Handoff prioritisation schemes are widely adopted in order to reduce the drop probability of handoffs in relation to new arrivals. Guard channels, queuing schemes, and combinations of them are proposed to reserving resources exclusively for handoffs providing guarantees for the handoffs dropping probability. Other research works that have dealt with the RRM issue include [24–26], while others have proposed CRRM techniques [27,28].

3. Resource management 3.1. Resource management plane In the initial provisioning phase, the network operator of the core network determines the maximum amount of bandwidth that can be dedicated on each link to each traffic class. It is the responsibility of the Admission Control (AC) mechanism to ensure that the traffic inserted in the network by each access network does not exceed such limits under normal network operation. We regard that each ER has an associated Admission Control Agent (ACA) which handles AC requests. Any application QoS requirement can be mapped into a reservation request for a set of traffic class parameters. The ACA is responsible for computing a single parameter, which is the bandwidth, out of the traffic class parameters based on the admission control algorithm [29] implemented for each traffic class. Therefore, Resource Pools are only aware of bandwidth as traffic parameter, while the AC can handle the whole set of parameters declared by a reservation request. However, the initial provisioning phase is not able to dynamically track fluctuations and deviations of the actual offered traffic from the expected one. In order to gain a more dynamic behaviour, it would be desirable to have a dynamic sharing of resources between different access networks, respectively, different ERs to some extent. That implies that the initial setting of each ER with a set of Admission Control Limits (ACLs), which depends on the offered traffic distribution, should dynamically adapt to the real offered load. Resource Pools comprise an original mechanism proposed in Ref. [6] to cope with local fluctuations in offered traffic, providing short-term resource re-distribution at a timescale of seconds to tens of minutes.

735

Resource Pools do not provide a centralized management approach, but introduce a distributed and scalable resource management framework. The concept of RP arises from the general consideration that there are constraints between the AC Limits of different ERs due to the topology and to the traffic distribution, and related to the presence of bottlenecks. The core idea is to organise adjacent ERs, which share a common bottleneck into groups (called Resource Pool) that can dynamically exchange their resources, i.e. co-ordinately increase/decrease their AC Limits, targeting at preserving scalability and implementation simplicity. An extended analysis about the algorithms related to the adaptation of the AC limits can be found in Ref. [30]. Application of the RP concept is straightforward in the case a set of ERs are connected to a Core Router (CR) following a star topology (Fig. 3), attached to the bottleneck. The above approach can be hierarchically extended to build RPs whose elements are not ERs but other RPs. In this way, a Resource Pool Tree (RPT) is created. Each leaf of the RPT, called Resource Pool Leaf (RPL), is associated to one Admission Control Agent. We have to stress here that the exchange of resources between the RP and the RPLs should follow some rules and guidelines in order to preserve scalability and simplicity. Different algorithms were developed and investigated in the framework of AQUILA [6], which covered issues like the amount of exchanged resources, the frequency of requests/release of resources, in order to achieve a high level of network utilization and low signalling load. In general, Resource Pools are based on the notion of ‘resource cushions’: unused resources are not immediately released to the above level RP, but kept for a period in order to be used for future reservations.

Fig. 3. How the RP notion is adopted.

736


3.2. RP algorithm for heterogeneous networks This section presents the Resource Management Algorithm (RMA) used in the RP to exchange resources between the root and the leaves, RPLs. The RMA presented here is based on watermarks on the amount of locally used resources (bandwidth): RPLs use a high and a low watermark to exchange resources with its RP. Resource share adaptation actions are triggered, when the measured total resource demand leaves the resource window enclosed between the two watermarks. A RPL sends a resource request to its RP, when total resource demand exceeds the high watermark. A RPL returns unused resources, when total resource demand falls below the low watermark. A change in the resource distribution is always initiated by a leaf of a Resource Pool Tree (RPT), when an AC request arrives that cannot be accepted because of lack of resources. This can cause a chain of resource requests running up a RPT. Notice that in view of a RPL there are two different resource requests: a primary request asking for resources of a RPL and possibly a triggered secondary resource request that is sent from the RPL to its RP in order to increase its resource share. The resource adaptation process of the watermark RMA is depicted in Fig. 5. Let r be the size of the resource share that is assigned to a RPL. A high watermark wh2[0,1] defines the threshold whr that is used to pass the primary resource request to the RP whenever the threshold is reached or crossed by the total resource demand. Before passing a primary resource request to a RP its resource demand is multiplied by nreq. A low watermark wl2[0,1] defines a second threshold w1r that is used to return the fraction nre1 of the unused resources above the low watermark to the RP whenever it is reached or crossed by the total resource demand. In order to detain unused resources for future requests, free resources are not released immediately to the RP, but detained for a time period defined implicitly by the counter cre1. Altogether the RMA uses the following variables and parameters and works as follows: Variables: r size of resource share that is assigned to considered RPL u amount of resources of RPL that is already reserved: 0%u%r q resource demand of considered primary resource request t actual time Parameters: wh2[0,1] high watermark expressed as fraction of the assigned resource share r nreqZ1,2,3,. multiplier for primary demand q of resource requests that is used by the RPL when it forwards a resource request that cannot be satisfied to its RP

low watermark expressed as fraction of the assigned resource share r nrel2[0,1] fraction of unused resources above low watermark that is returned when low watermark is crossed crelZ1,2,3,. primary request/release counter which triggers the resource release check RMA actions: For each primary resource request demanding q resource units check if total demand exceeds the high watermark: if uCqRwhr then request an additional block of resources of size of q$nreq from RP. If the sum of primary resource requests and releases exceeds the crel check if total demand falls below the low watermark: if u%wlr then return a block of unused resources of size of (1-wl)r$nrel to RP.

wl2[0,1]

3.3. Extended RP algorithm for mobility support The above description of the RMA algorithm does not differentiate whether the primary requests are due to new call arrivals or handoff requests. An extension of the RMA, namely the RMA-HF, provides handling differentiation between handoff and new requests. In case a handoff request cannot be accommodated neither by the RPL nor the RP, the RP will issue an urgent release request, requesting explicitly from all its other RPLs to release their free resources. The father RP will satisfy the request of the RPL, assigning to it the amount of resources to accept the handoff request, but will keep the remaining amount of free resources to redistribute them between its children. This scheme is in contrast to the handling of a new request, which will be rejected if there are no free resources at both the corresponding RPL and the RP.

4. Resource pools performance The performance of RPs was investigated with a simulation model that was developed using Java. Two performance investigations are presented. First RMA performance is systematically studied with a large number of simulations covering almost the total parameter space. It is shown that a RP can adapt resource shares and balances oscillating load efficiently. At the second part of simulations, a comparative study between the two different algorithms, RMA and RMA-HF is presented. Moreover, in order to clearly state the added-value of the RP mechanism, the performance of a static resource provisioning scenario under the same traffic conditions is presented. For a systematic performance investigation covering almost the whole parameter space two RPLs sharing a common pool were used as shown in Fig. 4. A traffic


r1 RPL 1

TG 1

R

λ1/µ1

RP

r2 TG 2

RPL 2

λ2/µ2 Fig. 4. Simulation model.

generator is connected to each RPL that generates primary requests with exponential distributed inter-arrival times and exponential distributed holding times. Requests are admitted as long as they do not require more resources that are free in the resource share of the requested RPL at that moment. For sake of simplicity, we assume that each reservation requests asks for one unit of bandwidth. Oscillating load with a mean of 80 flows respectively 40 flows was offered to RPL1 respectively RPL2 (see Fig. 5). RPLs run the RMA to adapt their resource shares to real resource demand. RP performance depends on resource provisioning policy, since available resources determine blocking probability. The RP started with 130 resource units in the common pool and 0 resources at the RPLs. This are 12% less resources than are needed to meet the target blocking probability of 1% with a static resource assignments to the AC functions of the RPLs according to the Erlang-B formula. Fig. 5 gives a detailed view of the behaviour of the resource adaptation process, where resources are exchanged between RPL1 and RPL2 continuously following load oscillations well. Resource shares follow resource demands at a coarser time scale wrapping up erratic demands with a smoother step function. Four performance measures were used to rate the performance of RMA: blocking frequency, fairness, adaptation rate, and capacity coefficient. Blocking frequency measures number of blocked primary resource requests accumulated over all RPLs in total. Individual blocking frequencies of the RPLs are used to measure RPL1-r1

RPL1-u1

RPL2-r2

737

fairness. The algorithm is fair, if individual blocking probabilities equal global blocking probability and are all the same. Instead of the adaptation rate itself, the ratio of secondary to primary reservation activity was measured per RPL (number of secondary resource requests from an RPL divided by number of primary resource requests at an RPL). The capacity coefficient is the ratio of the pool size to the amount of resources needed to get the same blocking frequency with static resource assignments to all RPLs. Figs. 6 and 7 show the results of the systematic analysis of RMA against different set of low watermark parameters. The low watermark determines willingness to share resources. Willingness to share decreases with lower low watermark values. Lower low watermarks are crossed less frequently thus freezing bonded resources. This has a large impact on performance as can be seen in Figs. 6 and 7.

RPL2-u2

120 110

Number of flows

100 90 80 70 60 50 40 30 20 10 0 2400

2900

3400

3900

4400

Time in secs

Fig. 5. Detailed view of the resource adaptation process.

Fig. 6. Measured impact of low watermark (w1, w2) on blocking frequency, activity and utilization of RPL1. The top picture shows severe performance at RPL1 when small values are used for the low watermark of RPL2 (w2%0.3). The picture in the middle shows increase of the resource adaptation rates, when large values are taken for low watermarks. Parameter setting: whZ1.0, nreqZ1, crelZ2, wl: w1, w22[0.10, 0.95], nrelZ0.8.

738


Fig. 7. Measured impact of low watermark (w1, w2) on blocking frequency, activity and utilization of RPL2. The top picture shows severe performance at RPL2 when small values are used for the low watermark of RPL1 (w1%0.7). The picture in the middle shows increase of the resource adaptation rates, when large values are taken for low watermarks. Parameter setting: whZ1.0, nreqZ1, crelZ1, wl: w1, w22[0.10, 0.95], nrelZ0.8.

There is a trade-off between RP activity and its performance. Best performance in view of a single RPL is obtained with high low watermark values at competing RPLs. But this leads to high low watermarks for all RPLs and high activity rates, see bottom picture of Figs. 6 and 7. Altogether there is an area in the front right quadrant of the low watermark plane only which results in fair resource sharing and low activity rates. The RMA is much less sensitive to the variation of the other parameters. Release block size has almost no influence, while the impact of request block size is larger but still very small compared to low watermark (see Ref. [11]). Variation of high watermark wh and release counter crel are not shown here because of lack of space. If high watermark is set to 1.0, additional resources will be requested if actually needed only. With high watermarks wh!1.0 resource shares are increased in advance. This is an

advantage if processing delay for secondary resource requests is high. Otherwise high watermarks should be as high as possible, because lower high watermarks increase activity. Inhibiting resource release for a number of reservation/release requests crel inflates for a short time the resource cushion and detains resources to be used for future requests. Increasing crel decreases the activity rate, while the blocking probability gets higher. As a second step of the performance analysis, the RMA was compared to the RMA-HF as well as to a static provisioning scenario. We regard that the ratio of new arrivals to handoff requests is 0.5, while special treatment to handoff calls is given only under the RMA-HF scheme. Based on the above analysis, there is only a small area in the front right quadrant of the low watermark plane, which results in fair resource sharing and low activity rates. Therefore, the low watermarks for RPL1 and RPL2 were set, respectively, 0.75 and 0.50. Moreover, in order to clearly state the difference of having high watermark values, Table 2 also presents the algorithms performance when both RPLs have a low watermark value of 0.95. Both algorithms worked well with the parameters setting of Table 1 in respect to all performance measures and against the static provisioning scenario (ST), as presented in Table 2. It is worth noting that the activity ratio is considerably affected by the increase of the low watermark values, while the blocking frequency is less sensitive, since both set of low watermark values is positioned within the area dictated by the above analysis. Therefore, the setting of parameters heavily depends on the performance target, e.g. in case activity rate does not comprise a great concern, then the values of low watermarks should get really high values. Fig. 8 depicts the distribution of the blocking frequency between different kind of requests—new ones and handoff ones—and different algorithms (low watermarks are set for both RPLs at the value of 0.95 in this analysis). It is obvious that the RMA-HF favours the handoff requests at expense of new arrivals. While under RMA both kind of flows share almost the same blocking frequency, 3.38%, under RMAHF handoff and new arrival requests have a blocking frequency of 2.1 and 4.3%, respectively. Moreover, it is worth mentioning the impact of handoff requests on new arrivals under the RMA-HF scheme, while the percentage of handoff calls increases. Under a percentage of 10% for the handoff requests, the blocking frequencies are 0.83 and 3.67% for the handoffs and the new Table 1 Parameter setting

wh nreq crel w1l w2l nrel

RPL1

RPL2

1 1 2 0.95 0.75 0.80

1 1 1 0.95 0.50 0.80


739

Table 2 Performance measurements results for RMA, RMA-HF and static provisioning (ST)

Blocking frequency% Fairness % Capacity coefficient Utilization Activity rate %

RMA-w1l

RMA-w2l

RMA-HF-w1l

RMA-HF-w2l

ST

3.38 b1Z3.51 b2Z3.38 gZ0.93 91.4 31.84

5.71 b1Z6.09 b2Z4.95 gZ0.98 85.8 8.78

3.17 b1Z3.05 b2Z3.43 gZ0.92 91.5 32.61

3.54 b1Z3.35 b2Z4.01 gZ0.93 89.6 13.19

7.69 b1Z8.19 b2Z6.71 – 83.8 –

arrivals correspondingly. When this percentage gets the value of 90%, the corresponding blocking frequencies are 2.7 and 4.7%.

5. Conclusions and future work This article briefly discussed the end-to-end QoS provisioning process within the heterogeneous next-generation wireless network environment. It presented the emerging next-generation network architecture and highlighted some important issues regarding QoS provisioning within this architecture. The main contribution of the paper deals with a dynamic resource management scheme for the Core Network of the next-generation wireless network architecture. The paper provided justifications on the need for such scheme and then analytically illustrated its operation, and its major performance parameters. A simulation model has been developed and results are presented in order to prove the appropriateness and correctness of the scheme. The scheme supports an ondemand redistribution of resources, in order to take care of prioritised situations, like that occurred during a handoff.

Simulations showed that the RPs achieve blocking probability targets with reasonable resources and distribute resources in a fair manner. The gain of RP is approximately 90% in terms of adaptation activity rate, meaning that only the 10% of primary resource reservation requests trigger an adaptation action. One possible extension of the proposed dynamic resource management scheme is to cater for precautionary measures in cases of handoff. Although the scheme already supports prioritised treatment of handoff resource needs, a possible further enhancement is to take into account the prediction of user movement, based on the notion of ‘mobile’ network services as well as on the appropriate mobility prediction algorithms. This would allow preallocating a specific portion of resources to the RPs that are going to need them in the near future with a high probability. Such an a priori allocation of resources would allow for less signalling traffic during a handoff, compared to the current situation, which implies that a handoff may cause an on-demand re-allocation of resources. However, we have to study the trade-off between the activity gain and the possible overall utilisation aggravation due to the preallocation of resources at an RP.

New Requests

Handoff

20 18

Blocking freguency %

16 14 12 10 8 6 4 2 0 ST

RMA

RMA-HF

RPL1

ST

RMA

RPL2

RMA-HF

ST

RMA

RMA-HF

Overall

Fig. 8. Blocking frequency % of new arrival and handoffs requests under different resource management schemes.

740


Another important issue for future work would be the extension of the application of the RP concept within the Access Network. In other words, one could extent the Resource Pool Leaf at e.g. the cell level. In this way, one could expect an overall superior management of the network resources, although interworking with radio resource management mechanisms would be inevitable. However, even though we may expect an increased complexity of the resource management scheme, this extension may alleviate the common RRM mechanisms that will be most probably employed in next-generation wireless network architectures.

References [1] GPP TS 23.234 v6.2.0, 3GPP system to Wireless Local Area Network (WLAN) Interworking; System Description (Release 6), September 2004. [2] K. Ahmavaara, et al., Interworking architecture between 3GPP and WLAN systems, IEEE Communications Magazine 41 (11) (2003) 74–81. [3] A. Salkintzis, Interworking techniques and architectures for WLAN/3G integration toward 4G mobile data networks, IEEE Wireless Communications 11 (3) (2004) 50–61. [4] The Internet Engineering Task Force, URL:http://www.ietf.org. [5] Internet2 QoS Working Group, URL:http://qos.internet2.edu/wg/. [6] AQUILA site page, URL:http://www-st.inf.tu-dresden.de/aquila/. [7] J. Gozdecki, Quality of service terminology in IP networks, IEEE Communications Magazine 41 (3) (2003) 153–159. [8] J. Wroclawski, Specification of the Controlled-Load Network Element Service, RFC 2211, September 1997 [9] S. Shenker et al.Specification of Guaranteed Quality of Service, RFC 2212, September 1997. [10] GPP TS23.207 v.6.4.0, End-to-end Quality of Service (QoS) concept and architecture, Release 5, September 2004. [11] N. Seitz, ITU-T QoS standards for IP-based networks, IEEE Communications Magazine 41 (6) (2003) 82–89. [12] A mobile differentiated services QoS model, Elsevier Computer Communications 27 (11) (2004) 1106–1114. [13] M. Islam, M. Murshed, Novel velocity and call duration support for Qos provision in mobile wireless networks, IEEE Wireless Communications 11 (5) (2004) 22–30. [14] X. Gao, End-to-end QoS provisioning in mobile heterogeneous networks, IEEE Wireless Communications 11 (3) (2004) 24–34. [15] N. Nafisi, et al., Extending QoS Policy-based mechanisms to B3G Mobile Access Networks, IST Mobile Summit, Lyon 27–30 (2004) 2004. [16] P. Demestichas, V. Stavroulaki, Issues In introducing resource brokerage functionality in B3G composite radio environments, IEEE Wireless Communications 11 (10) (2004) 32–40. [17] R. Bless, et al., Quality-of-service signaling for next generation IPbased mobile networks, IEEE Communications Magazine 42 (6) (2004) 72–77. [18] Y. Cheng, W. Zhuang, Diffserv resource allocation for fast handoff in wireless mobile internet, IEEE Communications Magazine 40 (5) (2002) 130–136. [19] V. Pandey, Exploiting user profiles to support differentiated services in next-generation wireless networks, IEEE Network 18 (5) (2004) 40–48. [20] G. Fodor, G. Fodor, et al., Providing quality of service in always best connected networks, IEEE Communications Magazine 41 (6) (2003) 154–163.

[21] S.-C. Lo, Architecture for mobility and QoS support in all-IP wireless networks, IEEE Journal on Selected Areas in Communications 22 (4) (2004) 691–705. [22] Next Steps in Signalling (NSIS) charter, http://www.ietf.org/html. charters/nsis-charter.html. [23] N. Tripathi, Handoff In cellular systems, IEEE Personal Communications 5 (6) (1998) 26–37. [24] A. Salkintzis, Radio resource management in cellular digital packet data networks, IEEE Personal Communications 6 (6) (1999) 28–36. [25] O. Sallent, An emulator framework for a new radio resource management for QoS guaranteed services in W-CDMA systems, IEEE Journal on Selected Areas in Communications 19 (10) (2001) 1893–1904. [26] A. Safwat, ACA: Channel assignment in ad hoc, 4G and beyond wireless networks with directional antennas, ICC 27 (1) (2004) 3143–3147. [27] A. To¨lli, Performance evaluation of common radio resource management (CRRM), ICC 25 (1) (2002) 3429–3433. [28] S. Uskela, Key concepts for evolution toward beyond 3G networks, IEEE Wireless Communications 10 (1) (2003) 43–48. [29] Deliverable D1302, Specification of traffic handling for the second trial, AQUILA project consortium. [30] T. Engel, et al., Analysis of adaptive resource distribution algorithms in the framework of a dynamic diffserv IP network ComCon 8 (8th International Conference on Advances in Communications and Control), Crete, Greece 2001.

Sotiris Maniatis received a Dipl.-Ing degree in computer engineering and informatics from the University of Patras, Greece, in 1996, an MSc degree in information systems engineering from UMIST, United Kingdom, in 1997, and a PhD degree from the School of Electrical and Computer Engineering of the National Technical University of Athens (NTUA) in 2002. He is currently a senior research associate in the ICBNet laboratory of NTUA. His primary research interests are mobile computing, mobility, QoS and resource management, and 3G and b3G mobile networks. He has received an academic excellence award from the National Scholarship Foundation and is a member of the Technical Chamber of Greece.

Eugenia Nikolouzou received her Dipl.-Ing and PhD degrees from the School of Electrical and Computer Engineering of NTUA, Greece, in 1999 and 2003, respectively. She is currently a senior research associate in the ICBNet laboratory of NTUA. Her research interests include QoS in IP networks, traffic engineering, resource management, and performance analysis. She has participated in various European Union and national projects. She is a member of the Technical Chamber of Greece.

Iakovos S. Venieris received a Dipl. -Ing. degree from the University of Patras, Greece, in 1988, and a PhD degree from NTUA in 1990, all in electrical and computer engineering. From 1992 to 1994 he was a research associate in the Telecommunications Laboratory of NTUA. In 1994 he became an assistant professor in the Electrical and Computer Engineering Department of NTUA, where he is now a Professor. His research interests are in the fields of broadband communications, Internet, mobile networks, Intelligent Networks, internetworking, signaling, service creation and control, distributed processing, agents technology, and performance evaluation. He has over 100 publications in the above areas. He has participated in several European Union and national projects. He is an associate editor of IEEE Communication Letters and a member of the editorial board of Computer Communications (Elsevier), and has been a guest editor for IEEE Communications Magazine. He is a reviewer for several journals and has been member of the Technical Program Committee and session chairman of several international conferences. He is a member of the Technical Chamber of Greece.