Dynamic admission control and bandwidth reservation for IEEE e mobile WiMAX networks

Transcription

1 RESEARCH Open Access Dynamic admission control and bandwid reservation for IEEE e mobile WiMAX networks Chiapin Wang *, Wan-Jhen Yan and Hao-Kai Lo Abstract The article presents a dynamic connection admission control (CAC) and bandwid reservation (BR) scheme for IEEE e Broadband Wireless Access networks to simultaneously improve e utilization efficiency of network resources and guarantee QoS for admitted connections. The proposed CAC algorim dynamically determines e admission criteria according to network loads and adopts an adaptive QoS strategy to improve e utilization efficiency of network resources. After new or handoff connections enter e networks based on current admission criteria, e proposed adaptive BR scheme adjusts e amount of reserved bandwid for handoffs according to e arrival distributions of new and handoff connections in order to increase e admission opportunities of new connections and provide handoff QoS as well. We conduct simulations to compare e performance of our proposed CAC algorim and BR scheme wi at of oer approaches. The results illustrate at our approach can effectively improve e network efficiency in terms of granting more connections by as large as about 22% in comparison wi oer schemes, and can also guarantee adaptive QoS for admitted new and handoff connections. Keywords: IEEE e WMAN, connection admission control, bandwid reservation, resource allocation. 1. Introduction Broadband wireless access networks have rapidly been growing in ese years to support e increasing demands of wireless multimedia services, like streaming audio/video, Internet Protocol TV, and video conferencing. Mobile Worldwide Interoperability for Microwave Access (WiMAX), which has been standardized by IEEE e [1], is one of e most promising solutions to provide ubiquitous wireless access wi high data rates, high mobility, and wide coverage. The IEEE e Media Access Control (MAC) layer provides differential Quality of service (QoS) for various classes of scheduling services, which are Unsolicited Grant Service (UGS), Extended Real-Time Polling Service (ertps), Real-Time Polling Service (rtps), Non-real-time Polling Service (nrtps), and Best Effort (BE). Each scheduling class is associated wi a set of QoS parameters for quantifying its bandwid requirement, e.g., maximum/minimum data rates and maximum delays. The radio resources (i. * Correspondence: chiapin@ntnu.edu.tw Department of Applied Electronic Technology, National Taiwan Normal University, Taipei, Taiwan e., time slots and frequency spectrums) for different scheduling services are centrally controlled by e base station (BS). To provide QoS for data transmissions in WiMAX networks, BS generally applies a Connection Admission Control (CAC) scheme which determines wheer a new connection should be established according to e available network resources. Essentially, e effectiveness of CAC schemes can be critical to bo e performances of QoS for admitted connections and e utilization efficiency of network resources. However, e IEEE e standards do not specify how to implement CAC mechanisms and remain at as open issues. On e oer hand, a bandwid reservation (BR) mechanism is also important to e provisioning of QoS for some prioritized users like users in a handoff process. Handoff occurs when mobile station (MS) transfers its connection from e original serving BS wi worse and worse link qualities to a neighboring BS wi better qualities. In general, a handoff user will be prioritized over a new incoming user in order to provide better user-perceived satisfaction especially when it is wi real-time applications which have specific QoS 2012 Wang et al; licensee Springer. This is an Open Access article distributed under e terms of e Creative Commons Attribution License ( which permits unrestricted use, distribution, and reproduction in any medium, provided e original work is properly cited.

2 Page 2 of 20 requirements, e.g., roughput demands and delay/jitter constraints. Since e reserved bandwid cannot be taken by a new coming user, e design of BR mechanisms can significantly affect e performance of handoff QoS and also e utilization efficiency of network resources. The CAC and BR problems have largely been investigated in previous study [2-20]. The auors of [2,3] propose to adopt minimum bandwid requirements as e admission criteria for all classes of scheduling services. The approach can provide more connections admitted into networks but may cause a relatively low QoS performance. The auors of [10] propose to divide e scheduling services into two groups: one group consists of UGS, ertps, and rtps which adopt maximum bandwid requirements for e admission criteria, while anoer group consists of nrtps and BE which adopt minimum bandwid requirements. The approach may over favor e higher-class services and cause a starvation of lowerclass services. Instead of using fixed criteria for an admission control as described above, e studies in [11,12] propose to dynamically determine e admission criteria by using a game-eoretic approach. However, it does not take e network load into consideration and may introduce great computational complexities. Wi regard to e BR schemes, a fixed guard channel scheme [13] is proposed to reserve a certain amount of bandwid for upcoming handoff connections to assure seamless handoff processes. When e total bandwid utilization of existing users reaches e reshold, no more new connections can be admitted into e network. Nevereless, when a fixed amount of bandwid can never be used for new connections, a certain portion of network resources may be wasted. The study [10] proposes to dynamically adjust e quantity of reserved bandwid based on e arrival and departure behavior of handoff connections to make e resource utilization more efficient. However, if handoff connections occur infrequently, e quantity of reserved bandwid for handoffs is almost fixed and is approach would be similar to e fixed guard channel scheme and cause a waste of network resources as well. Bo CAC schemes and BR mechanisms are important research issues in wireless networks due to scarce radio resources, dynamic channel qualities, and diverse user demands. However, to e best of auors knowledge, most efforts tackle one of e two problems individually whilelittleworkconsidersejointdesignofetwo mechanisms. We are us motivated to present a joint design of CAC and BR mechanisms which aim at simultaneously improving e utilization efficiency of network resources and guaranteeing QoS for admitted new connections and handoff connections. The proposed CAC scheme dynamically determines e admission criteria according to network loads and adopts an adaptive QoS strategy to increase e amount of admitted connections for e network efficiency. The key idea of our CAC scheme is based on e fact at most scheduling services are wi adaptive QoS requirements, e.g., maximum and minimum rates. Therefore, e admission criteria can be determined according to e amount of available wireless resources for increasing e number of admitted connections wi adaptive QoS. For example, if e network capacity is adequate or sufficient, bandwid requirements for higher QoS might be adopted as e admission criteria. Alternatively, if e network load is quite heavy, e admission criteria may be degraded to meet lower QoS requirements. After e admission criteria are determined, e proposed BR scheme dynamically adjusts e amount of reserved bandwid for handoffs according to e arrival distributions of new/handoff connections to increase e connection admission opportunities and also guarantee e bandwid requirements for handoff QoS. The basic idea of our adaptive BR scheme is a rational inference at generally e occurrences of new incoming connections may be much more frequent an at of handoff connections [21-24]. This observation originates from common BS deployment at e overlap areas of a given BS between its neighboring stations are parts of its coverage area. Since handoffs arise only when users cross rough e overlap areas, it is a general situation to observe more new connections occurring an handoff connections. Thus, e optimal BR should take into account e arrival behavior of not only handoff connections, but also new connections in order to avoid a waste of network resource as possible. We conduct simulations of e transmission scenarios to evaluate and compare e performances of e proposed CAC algorim and BR scheme wi at of oer approaches. Simulations results illustrate at our approach can effectively improve e network efficiency in terms of increasing e number of granted connections by as large as about 22% in comparison wi oer schemes, and also can guarantee adaptive QoS for admitted new and handoff connections. The remainder of is article is organized as follow. In Section 2, we briefly illustrate e QoS architecture and resource allocation mechanism of IEEE e networks. Section 3 presents e proposed CAC algorim and BR scheme. In Section 4, we construct simulation scenarios to demonstrate e effectiveness of our approach. Section 5 draws our conclusions. 2. IEEE e QoS architecture and resource allocation mechanism 2.1 IEEE e QoS architecture The IEEE e MAC layer provides QoS differentiation for various categories of scheduling services. The

3 Page 3 of 20 IEEE802.16e uplink scheduling framework is shown in Figure 1. The scheduling of uplink packet transmissions is centrally controlled in e BS. The IEEE e standards adopt a connection-oriented MAC protocol, i.e., each connection is associated wi a connection ID. When a service flow generated at e application layer arrives at e MAC layer, e MS first sends a connection establishment request to e BS. The admission control mechanism at BS en estimates wheer e remaining bandwid can support e QoS requirements of new connections wiout violating existing users QoS. If e connection request is accepted, e BS replies wi a connection response which indicates e connection IDs for each direction of is connection. After e process of connection establishment is finished, e MS can issue a bandwid request. The connection classifier en classifies e service data units into different scheduling classes according to eir service flow identifier and connection identifier. The uplink bandwid requests by users are performed on a per connection basis, whereas e BS grants bandwid on a per subscriber station basis (GPSS). After e BS allocates a certain amount of bandwid to each of e MSs, e packet scheduler at each MS will redistribute e bandwid to e corresponding connection. By means of e connection-admission-control mechanism and request-grant bandwid-allocation scheme, QoS for different scheduling classes can be guaranteed. The IEEE e standard divides all service flows into five scheduling classes, each of which is associated wi a set of QoS parameters for quantifying its bandwid requirement. The five scheduling classes are described as follows. (1) UGS: UGS is designed to support real-time service flows wi fixed-size packets generated at periodic intervals (i.e., constant bit rate CBR), such as T1 services and voice-over-internet-protocol (VoIP) applications wiout silence suppression. This service can grant a fixed amount of bandwid for CBR real-time applications wiout any requests. (2) rtps: rtps is designed to support real-time service flows wi variable-size packets generated at periodic intervals (i.e., variable bit rate VBR), such as Motion Pictures Experts Group (MPEG) video. Based on a polling mechanism to request bandwid periodically, is service can guarantee QoS such as e minimum data rate and maximum latency for VBR real-time applications. (3) ertps: The characteristic of is service class is between UGS and rtps. On detecting at e allocated bandwid is eier insufficient or excessive, ertps can send a request to change e amount of allocated bandwid like rtps does. Oerwise, if e bandwid demand remains unchanged, ertpsbehavesasugs. ertps is designed to support VBR real-time data services such as VoIP applications wi silence suppression. Figure 1 IEEE802.16e uplink scheduling framework.

4 Page 4 of 20 (4) nrtps: This service class is to support non-realtime VBR services which require minimum-data-rate guarantees but can be tolerant to delay, such as File- Transfer-Protocol (FTP) applications. (5) BE: The BE service is designed for best-effort applications which have no explicit QoS requirements, e.g., web services or . The QoS parameters and e supporting application types associated wi each of e IEEE e scheduling classes are shown in Table Bandwid allocation mechanism The IEEE e physical layer (PHY) adopts an Orogonal Frequency Division Multiple Access (OFDMA) slot as e minimum possible resource. The IEEE e PHY supports Frequency Division Duplex (FDD) and Time Division Duplex (TDD) for bandwid allocation mechanisms. In FDD mode, e uplink (UL) and downlink (DL) channels are located on split frequencies, wi which a fixed duration frame is used for bo UL and DL transmissions. In TDD mode, e UL and DL transmissions are arranged at different time periods using e same frequency. In is article, we focus on e TDD mode for e IEEE e bandwid allocation mechanism. In TDD mode, Time Division Multiplexing (TDM) is used for DL transmissions and Time Division Multiple Access (TDMA) is used for UL transmissions. As shown in Figure 2, a TDD frame has a fixed duration and contains one DL subframe and one UL subframe whose durations can adapt to e traffic loads of UL and DL transmissions. The DL subframe consists of a preamble, Frame Control Header (FCH), and a number of data bursts. The FCH specifies e profiles of e DL bursts at immediately follow it. The broadcast messages Table 1 IEEE802.16e QoS classes QoS classes UGS Applications T1 services, VOIP wiout silence suppression QoS parameters Max Rate Min Rate Jitter rtps Video Streaming Max Rate Min Rate Max Latency ertps VOIP wi silence suppression Max Rate Min Rate Max Latency Jitter nrtps FTP Max Rate Min Rate BE Web browsing, Max Rate including downlink map (DL-MAP), uplink map (UL- MAP), DL Channel Descriptor (DCD), UL Channel Descriptor (UCD), etc., are sent at e beginning of ese DL bursts. The UL subframe contains a contention interval for initial ranging and bandwid request andulphyprotocoldataunits(pdus)fromdifferent MSs. The DL connections are scheduled by BS in a broadcast manner, while e UL connections apply a request-grant mechanism for bandwid allocation in a shared manner. The UL bandwid requests are performed on a per connection basis, whereas e BS grants bandwid on a per subscriber station basis (GPSS). After e BS allocates a certain amount of bandwid to each of e MSs, each MS will redistribute e bandwid to e corresponding connection. The information about bandwid allocations for DL and UL transmissions is broadcast to e MSs rough DL-MAP and UL-MAP messages at e beginning of each frame. Therefore, each MS can receive from and transmit data to BS in e predefined OFDMA slots Packet scheduling mechanism As shown in Section 2.1, e IEEE e standard defines five scheduling classes. However, it does not specify e scheduling mechanism for e five classes and e design is left for researchers [25]. The design of a scheduling mechanism must take into account e specific QoS constraints of different applications, e.g. e maximum allowable delay and minimum data rate [3]. A feasible solution is to decide on a service class first according to e characteristics of each class and next choose an appropriate user in e selected class [26]. In e second phase, e packet scheduling of different users among a given class may consider some performance metrics such as roughput and fairness, while e maximum rate scheduling (greedy algorim) and Proportional Fairness (PF) scheduling can be applied, respectively. The maximum rate scheduling is effective to advance e overall system roughput as it allocates resources to users wi relatively good channel qualities among em [27]. On e oer hand, e PF scheduling can improve e fairness of channel utilization among users as it distributes resources among em wi consideration of eir previous records of utilization [28-30]. Throughput and fairness, however, are conflicting performance metrics [31]. To maximize system roughput, more resources should be allocated to e users in good channel conditions. This may cause most radio resources monopolized by a small number of users, leading to unfairness. On e contrary, if resources are allocated in a fair manner, resources may be allocated to e users wi weak channel conditions. This can result in e degradation of system roughput. To escape

5 Page 5 of 20 DL Subframe DL-PHY PDU Contentioninitial ranging Frame UL Subframe Contention bandwid request UL-PHY PDU from SS #1 UL-PHY PDU from SS #2 Preamble FCH DL burst #1 DL burst #2 Preamble UL burst DLFP DL-MAP, UL-MAP, DCD, UCD MAC PDUs MAC PDUs MAC PDUs PAD MAC Header MAC Payload CRC Figure 2 The TDD frame architecture (Source: IEEE [1]). from e roughput-fairness dilemma, we can consider utility for packet scheduling. Utilities are a performance metric which can fully represent e degree of user satisfaction for a given application [32,33]. Thus, it is a more appropriate metric for packet scheduling since even users wi e same service class actually are wi various demands for network resources due to eir specific characteristics. This fact implies at resources should be allocated to users according to e application performance metric of satisfaction raer an network performance metrics such as roughput or fairness [34]. In is article, we focus on e problems of CAC and BR for mobile WiMAX networks and do not discuss e design of packet scheduling. In e following section, we will propose a joint design of CAC and BR mechanisms as a solution of e addressed problems. 3. CAC algorim and BR scheme 3.1. Dynamic CAC algorim We design a dynamic CAC algorim which adjusts e admission criteria depending on e network loads and uses an adaptive QoS provisioning strategy in order to increase e efficiency of bandwid utilization. The key idea of our dynamic CAC scheme is based on e fact at most scheduling services are wi adaptive QoS requirements, e.g., maximum and minimum rates. Therefore, e admission criteria can be adjusted wi e amount of available wireless resources to increase e number of admitted connections wi adaptive QoS. For example, if e network capacity is adequate or sufficient, higher QoS requirements could be adopted as e admission criteria. Alternatively, if e network load is raer heavy, e admission criteria may level down for lower QoS. We can take into account several approaches to design e admission criteria according to network loads. A simple way, as shown in Figure 3a, is to adopt e maximum or minimum bandwid requirement alternatively as e admission criterion depending on e network load nl wi respect to a reshold, g. That is { bi,max, nl <γ b i = b i,min, nl γ, (1) where b i is e admission criteria for connection i; b i, max and b i, min are e maximum and minimum bandwid requirements corresponding to e highest and lowest QoS, respectively, for connection i. Furermore, we can consider a linear adaptation approach which smooly regulates e admission criteria b i in proportion to e change of network load nl asshowninfigure 3b. Accordingly, b i can be expressed as 1,nl < nl min nl max nl b i = αb i,max +(1 α)b i,min, whereα = nl max nl min 0,nl nl max, nl min nl < nl max where nl min and nl max refers to e minimum and maximum reshold of network loads respectively. If nl is less an nl min, e weighted factor a will be set as 1 and b i,max would be adopted as e admission criterion to provide e best QoS. On e oer hand, if nl is equal to or larger an nl max, a will be set as 0 and b i, min wouldbeusedtoprovideelowestqos.ifnl is (2)

6 Page 6 of 20 Admission criteria, b i b i,max b i,min a. Hard-decision function γ network loads Admission criteria, b i b i,max b slope = nl i,min max b i,max nl min b i,min nl min max nl network loads b. Linear function 0 b i 1 b i 2 b i 3 b i b i,max N 2 b i N 1 b i N b i b i,min nl nl nl N 2 nl N 1 nl c. Quantized-step function Figure 3 Several functions to adjust e admission criteria according to e network load. (a) Hard-decision function. (b) Linear function. (c) Quantized-step function. wiin e range of [ nl min, nlmax ), b i will be linearly regulated in proportion to e change of nl to provide an adequate QoS. In general, nl min and nl max can be varied wi e network bandwid B and each type of service. Wi a given B, nl min and nl max for a higher service class should be larger in order to provide differential QoS among various service classes. A feasible approach to determine e values of nl min and nl max

7 Page 7 of 20 wi regard to each service class is to set different ratios of nl min and nl max to B appropriately. For example, in e simulations we consider e following setting of ( nl min /B, nlmax /B): (3/6, 5/6), (2/5, 4/5), (2/6, 4/6), (1/5, 3/5), and (1/6, 3/6), respectively, for UGS, rtps, ertps, nrtps, and BE in order. Finally, as shown in Figure 3c, is approach is to quantize e bandwid wiin [b i, min, b i,max ]win + 1 levels, each of which refers to a feasible admission criteria. The adaptive criteria b i can be expressed as b 0 i (b i,max), nl < nl 0 (nlmin ) b i = b n i, nln 1 nl < nl n,1 n N 1 (3) b N i (b i,min ), nl nl N 1 (nl max ) where nl n refers to e n reshold of network load (1 n N -2).Inparticular, nl 0 and nln 1 refer to e minimum and maximum resholds of network loads, nl min and nl max, respectively. Here, b k i (1 k N -1)referstoek feasible admission criteria as e network load nl is wiin e range of [ nl k 1, nlk ). In particular, b 0 i corresponds to b i, max as nl < nl 0, and bn i corresponds to b i, min as nl nl N 1, respectively. The value of N + 1 refers to e number of quantization levels. When N is 1 in particular, e quantized step function has two levels and is equal to e hard-decision function as shown in Figure 3a. In contrast, if N is infinite, e quantized step function has countless levels and is equal to e linear adaptation approach as shown in Figure 3b. Normally, e value of N can be determined considering a reasonable number of adaptive criteria b i between b i, min and b i, max,(e.g.,n =5-15in general). When e value of N is determined, e unknown values of nl n (1 n N -2)andbk i (1 k N -1)canereforebeobtainedbyusingauniform quantizer which equally partitions e region wiin ( b 0 i, b N i )and(nl 0, nln 1 ). In is article, we consider e linear adaptation approach as shown in Figure 3b for our CAC algorim and use it for performance evaluations in e following section. Denote e bandwid allocated to existing new connections and handoff connections as b n and b h, respectively. A handoff connection will be accepted in e networkaslongasebandwid available can satisfy e bandwid requirement. Thus, e condition to accept a handoff connection is ho accepted = (b i,ho + b n + b h ) B, (4) where b i,ho is e admission criterion of handoff connection i determined in Equation (2). A new connection will be accepted when e following condition is satisfied: new accepted = ((b i,new + b n ) ad ) ((b i,new + b n + b h ) max ), (5) where b i, new is e admission criterion of new connection i provided in Equation (2). The first term in e rhs of Equation (5) aims for increasing e admitting opportunities for new connections when it only determines wheer e resources allocated to new connections exceed e reshold. The last term in e rhs aims for ensuring a minimum BR for handoff (i.e. B - max ). In Section 3.3 we will have a more detailed illustration for e criteria considered in Equation (5) Estimation of system capacity The system capacity B may be dynamic as e IEEE e standards on PHY support multiple transmission rates by using adaptive modulation and coding (AMC) schemes. The transmitter will determine one from various modulation and coding schemes (MCSs) available according to e channel conditions of packet delivery to provide reliable link qualities, large network coverage, and high data rates as possible. The modulation types supported in e IEEE e standards include Binary Phase-Shift Keying (BPSK), Quadrature Phase-Shift Keying (QPSK), 16-Quadrature Amplitude Modulation (16-QAM), and 64-QAM. Wi e Convolutional Turbo Code (CTC) and different code rates, e MCSs provided for WiMAX wi 5 and 10 MHz channels are summarized in Table 2[35]. In a DL transmission for example, each MS informs its current perceived channel quality to e BS periodically, and en e BS will choose a specific MCS corresponding to is channel condition. The transmission data rate wi a given MCS can be evaluated as R MCSi =(n Data SC /T S ) b MCSi, (6) where n Data_SC is e number data sub-carriers; T S is symbol period, and b MCSi is e amount of information bits per symbol wi respect to e i MCS, MCS i.in addition, adopting a Multiple-Input and Multiple-Output (MIMO) mechanism can furer increase e transmission data rates to several-fold e original amount. Here, our CAC scheme estimates e current system capacity B by consideration of e proportion of used MCSs [36,37]. Take an example for e capacity estimation as follows. Consider a 10-MHz channel spectrum wi a 2 2 MIMO mechanism in e downlink transmission. Assume e proportion of used MCSs is QPSK 3/4 = 25%, 16-QAM 1/2 = 25%, and 64-QAM 5/6 = 50%. Thus, e estimated system bandwid B in is case will be (9.5*2)*25% + (12.67*2)*25% + (31.68*2) *50% = Mbps. In general, B in e downlink can

8 Page 8 of 20 Table 2 The mobile WiMAX PHY data rates (Source: WiMAX Forum [35]) Parameter Downlink Uplink Downlink Uplink System bandwid 5 MHz 10 MHz FFT size Null sub-carriers Pilot sub-carriers Data sub-carriers Symbol period μs Frame duration 5 ms OFDM symbols/frame 48 Data OFDM symbols 44 Mod Coding rate 5 MHz channel 10 MHz channel Downlink rate (Mbps) Uplink rate (Mbps) Downlink rate (Mbps) Uplink rate (Mbps) QPSK 1/2 CTC, 6x /2 CTC, 4x /2 CTC, 2x /2 CTC, 1x /4 CTC QAM 1/2 CTC /4 CTC QAM 1/2 CTC /3 CTC /4 CTC /6 CTC range from 2.12 to Mbps widely wi AMC in different channel situations. Based on e number of supported users wi respect to each MCS and e proportion of adopted MCSs, our CAC scheme can estimate e current system capacity and network load and update e information periodically every frame period in e BS site Adaptive BR scheme We propose an adaptive BR scheme which dynamically adjusts e amount of reserved bandwid for handoffs according to e arrival distributions of bo handoff and new connections. The objective of our scheme is to simultaneously increase e admission opportunities for new coming users and guarantee QoS for handoff users. The basic idea of our adaptive BR scheme is a rational inference at generally e occurrences of new incoming connections may be much more frequent an at of handoff connections [22-24]. Thus, e optimal BR should take into account e arrival behavior of not only handoff connections, but also new connections in order to avoid a waste of network resource as possible. The scheme considers two resholds for BRs, min and max, which refer to e minimum and maximum resholds of reserved bandwid, respectively. The adaptive reshold of BR, ad, is initially set as ( min + max )/2, and will be dynamically adjusted wiin e range [ min, max ] according to e arrival behavior of new connections and handoff connections to control e reserved bandwid for handoff connections. When e condition shown in Equation (4) is met and a handoff connection is accepted, e adaptive reshold ad will be decreased wi e amount of allocated resources for e handoff connection wiout going below min. That is ad = min ( ad b i,ho, min ) (7) Alternatively, when e condition shown in Equation (5) is met and a new connection is accepted, ad will be increased wi e amount of allocated resources for e new connection wiout exceeding max. That is ad =max( ad + b i,new, max ) (8) To more clearly illustrate e characteristics of our proposed BR scheme, we conduct a simplified transmission scenario to provide a preliminary performance comparison between our scheme and e fixed reshold (FT) and dynamic reshold (DT) schemes [10]. Wi e FT scheme, e reshold of reserved bandwid, fixed, is fixed; e new connection would be accepted

9 Page 9 of 20 only if (b i, new + b n + b h ) fixed. Wi e DT scheme, e reshold of reserved bandwid, dyn,willbe adjusted wiin e range [ min, max ]dependingon e arrival and departure of handoff connections as follows: when a handoff connection is accepted, e reshold dyn will be increased wi e amount of resources allocated to e handoff user; when an existing handoff connection terminates, dyn will be decreased wi e amount of released resources. The new connection will be accepted only when (b i,new + b n + b h ) dyn.wi regard to handoff connections, all e ree algorims have e same admission strategy at a handoff connection will be accepted as long as e amount of bandwid available can meet its requirement, i.e. (b i,ho + b n + b h ) B. We exploit e following simplified scenario to compare e performances of e ree BR schemes. Assume at e total amount of network resources B is 100 units. Consider at e network is empty in e beginning. Assume at 80 new connections, 5 handoff connections, and 5 new connections arrive sequentially. Assume at each of e arrival connections requests a unit of resources. For e FT scheme, fixed is set as 80 units. In e DT scheme, min and max are 80 and 90 units, respectively, and dyn is set as 80 units initially. In our scheme, min and max are 0 and 90 units, respectively, and consequently e initial value of ad is 45 units. Figure 4 shows e process of connection admissions wi e ree schemes, respectively. It is shown at wi e FT scheme, e first 85 connections (80 new and 5 handoff connections) are accepted. But, e last five new connections are rejected due to a fixed reservation manner at a certain amount of bandwid can never be used by new connections. The DT scheme has e same performance as e FT scheme at e last five new connections are blocked. The reason is at e DT scheme adjusts e reserved bandwid according to e arrival and departure behavior of handoff connections only. Thus, when e occurrence of handoff connections is relatively infrequent as e scenario shown above, e DT scheme will be similar to e FT scheme and can cause a waste of network resources as well. Wi our scheme, e 90 connections totally can be granted into e network. Note at our scheme can still be effective in e situations when e number of new connections is nearly equal to or less an at of handoff connections (is will be furer examined in e following section). It is shown at e proposed scheme outperforms e two schemes in terms of increasing e number of admitted connections. Alough e DT scheme and our proposed scheme bo apply an adaptive manner for eir BR strategies, ere are two essential differences which lead to eir dissimilar performances. One difference is between eir criteria for accepting new connections. For e DT scheme, e new connection will be accepted if e amount of its bandwid requirement and e bandwid allocated to existing connections would not exceed e reshold, i.e. (b i,new + b n + b h ) dyn. In our scheme, one criterion for accepting a new connection as shown in e first term of e rhs in Equation (5) only examines wheer e amount of e requirement and e bandwid allocated to existing new connections does not exceed e reshold, i.e. (b i,new + b n ) ad.thus,wi our scheme, e existing handoff connections will not lessen e resource for an incoming new connection. That is, when e DT scheme and our proposed scheme are wi e same amount of reserved bandwid ( dyn = ad ), our scheme can increase a resource space of b h for accepting more new connections. Meanwhile, our BR scheme can guarantee QoS for handoff connections because of anoer condition for accepting a new connection as shown in e last term of e rhs in Equation (5). That is, e total amount of required bandwid for a new connection and allocated bandwid for existing connections would not exceed e amount of guarded bandwid for handoff, max,i.e.(b i,new + b n + b h ) max. Thus, at last e B - max bandwid amount can be reserved for handoff users. Anoer difference is between eir adjustment strategies for e reshold of BR, dyn and ad.wiedt scheme, dyn is increased or decreased when a handoff connection is established or terminated respectively, and will remain constant wheer a new connection concludes or runs its course. When e occurrence of handoff connections is relatively infrequent as e scenario shown above, e DT scheme will be similar to e FT scheme and can cause a waste of network resources too. In our scheme, ad is increased or decreased when a new or handoff connection is admitted and runs its course, respectively, and will remain constant wheer a new or handoff connection concludes. Note at a lower reshold of BR advances e acceptance rate of handoff users whereas a higher reshold increases e admission opportunities for new connections. In e sense, our scheme has potential to grant more new connections into networks especially when e occurrences of new connections are much more an at of handoff connections. As aforementioned, our scheme would not sacrifice handoff QoS for favoring new connections when no less an e bandwid amount of B - max will be reserved for handoff users. Thus, e proposed adaptive BR scheme can simultaneously improve e network efficiency by granting more new connections and also guarantee handoff QoS. To summarize, Figure 5 presents e joint design of e proposed CAC algorim and BR scheme, and

10 Page 10 of 20 a. FT scheme b. DT scheme c. BR scheme Figure 4 The process of connection admissions wi e ree BR schemes, respectively. Figure 6 shows e corresponding pseudocodes. The block diagram in Figure 5 contains e following ree steps. (1) When a new connection or handoff connection arrives, it will inform e BS of its highest and lowest bandwid requirements (i.e., b i, max and b i, min ). The proposed dynamic CAC scheme will adjust e admission criterion using Equation (2) according to e currently estimated system capacity and network load. (2) When e admission criterion is determined, e proposed adaptive BR scheme will accept or reject is handoff or new connection depending on e criterion in Equation (4) or Equation (5), respectively. (3) If a handoff connection is established, ad will be decreased wi e amount of allocated resources as Equation (7) shows; if a new connection is granted, ad will be increased wi e amount of allocated resources as Equation (8) shows.

11 Page 11 of 20 QoS Requirements of Connection i (b i,max, b i,min ) Network Loads (1) (1) ad = ad + b i,new b i = α b i,max + (1-α) b i,min Admission Criteria (2) Accepted or rejected? Dynamic Connection Admission Control (CAC) Algorim (3) Accepted ad = ad - b i,ho Bandwid Reservation Threshold New New or Handoff? Handoff Adaptive Bandwid Reservation (BR) Scheme (3) Figure 5 The proposed dynamic CAC algorim and adaptive BR scheme. Figure 6 CAC and BR algorim. ad ( At time epoch t forall (pending connections c and service flow i of c) is if ( type is if ( b if ( b else endif endforall max in{ UGS, ertps, rtps, nrtps, BE} do b = b i i b = b i i, ho if ( b ad endif else // no more capacity in B if ( b endif i, ho ) / 2 max( i, new ad + i,max i,min i, new min if ( BS more BW elseif ( BS ordinary BW b = α b i,max else ( BS less BW n n reject new n accept new min( ) of class i + (1 α) b accept handover n h h ) of class i min, reject handover + b <= max ad ( ad,( ) i,min handover and service class is + b + b <= B) en + b + b <= ad max else // it is new connection and b ad ) en b i, ho i, new + b <= ) en + b of class i )) i, new )) n h i) en + b + b <= max ) en

12 Page 12 of 20 The implementation of e proposed CAC and BR schemes in practice can involve e overheads as follows. (1) The estimate of system capacity: The system capacity can be evaluated at e BS wi e specific PHY characteristics like channel spectrum, e amount of data sub-carriers, supported MCSs, used MIMO mechanisms, etc. The estimation of system capacity can be obtained in e initial phase of a network built-up. (2) The estimate of network loads: In general, e network loads can be evaluated at e BS wi e information of currently adopted MCSs and e number of supported users wi respect to each MCS. This part may need e exchanges of some context information between BS and SSs periodically, e.g., currently channel condition and used modulation. (3) The determination of admission criteria for incoming connections: When a connection arrives and requests for an admission, it will inform e BS of its specific QoS requirements, e.g., maximum and minimum data rates. Based on e estimated system capacity and network loads and QoS parameters, e BS will compute e admission criteria for e incoming connection wi its specific QoS parameters. (4) The adaptation of BR for handoff connections: Ifa connection is admitted in e network, e BS will erefore adapt e BR reshold depending on e type of connection, i.e., new or handoff. 4. Performance evaluations and results In is section, we conduct simulations of e transmission scenarios to demonstrate e effectiveness of e proposed CAC algorim and BR scheme. The simulator is constructed in C and followed e IEEE e standard closely [35,38]. The channel spectrum is 10 MHz. The MAC frame duration is 5 ms, which consists of 1024 OFDM subcarriers (840 data and pilot subcarriers). One MAC frame includes 48 OFDM symbols, while e first symbol is used for a preamble. The ratio of e symbols of e uplink subframe to ose of e downlink subframe is 18:29. In e uplink, ree symbols are used for control signaling, and ere are 44 OFDM symbols used for data transmissions in e uplink and downlink in total. The simulation set-up considers a 2 2 MIMO mechanism and e AMC schemes. We used e following distribution for MCS levels: QPSK 1/12 = 3.71%, QPSK 1/8 = 12.01%, QPSK 1/4 = 29.10%, QPSK 1/2 = 29.67%, QPSK 3/4 = 9.23%, 16-QAM 1/2 = 12.51%, 64-QAM 1/2 = 0.75%, and 64-QAM 3/4 = 3.02% [39,40]. The OFDMA PHY parameters and eir values are listed in Table 2. We provide different simulation scenarios to examine e proposed CAC algorim and BR scheme individually or jointly to clearly show e performance effects wi e two schemes. For each of e scenarios, we assume e connection arrivals and departures are wi e Poisson process wi a mean arrival rate l and a mean departure rate equal to one-ten of e arrival rate. We assume at e BS is aware of e amount of connections wi regard to each kind of MCS and e bandwid requirement of each connection. The BS can erefore exploit is knowledge to estimate e current system capacity and also network loads. Generally, e system capacity and network loads can erefore be estimated for our scheme wi e information of currently adopted MCSs and e number of supported users wi respect to each MCS. The total simulation period is 1000 s while e results are provided wi e average values over 20 times of simulations The connection blocking rates and achieved roughput wi different CAC schemes In is section, we examine e performance of our CAC algorim individually by e comparisons wi at of oer CAC schemes. Here, we do not take into account e reserved bandwid for handoff users. Thus, each of e incoming connections will be admitted into e network as long as e bandwid available can satisfy e admission criteria determined by various CAC schemes. The simulation set-up for is scenario considers ree types of traffic classes which are rtps, ertps, and nrtps. To clearly examine e achieved QoS performances regarding e ree types of service classes wi different schemes, we separately evaluate e performance for each class wi its specific QoS requirement, i.e., maximum and minimum data rates as shown in Table 3. We compare e performance of our CAC algorim wi at of e static maximum (Static-max), static minimum (Static-min), and bandwid adaptation (Adapt) scheme considered in [11,12]. The Static-max and Static-min schemes adopt e highest and lowest QoS requirements, respectively, as e admission criteria. The Adapt scheme considers e highest QoS criteria for a new coming user. If e resources available are insufficient to meet e bandwid requirement of a new connection, e bandwid allocated to existing Table 3 The maximum and minimum rates associated wi different scheduling services Scheduling services Maximum rate (kbps) Minimum rate (kbps) UGS rtps ertps nrtps BE 20 0

13 Page 13 of 20 users will be decreased to satisfy e QoS requirement of e new connection. In our scheme, we consider a linear adaptation approach as shown in Figure 3b which adjusts e admission criteria according to e variation of network loads. We use e following ratios of ( nl min /B, nlmax /B): (2/5, 4/5), (2/6, 4/6), and (1/5, 3/5), for rtps, ertps, and nrtps, respectively, as we consider in Section 3.1. Then, e minimum and maximum resholds of network loads ( nl min, nlmax )inequation (2) for rtps, ertps, and nrtps can be derived, respectively, depending on e currently estimated system capacity B. The performances are indexed as e connection blocking rate and e achieved QoS in terms of per-flow roughput, i.e., e average data rates supported per established connection. Figure 7 presents e average connection blocking rates of e rtps, ertps, and nrtps flows wi different CAC schemes. It is shown at e blocking rates rise wi e increase of connection arrival rates l for all schemes except e Adapt scheme. The blocking rate wi e Adapt scheme is almost 0 steadily, which is e lowest among all schemes when e incoming connections will always be established regardless of e amount of bandwid available. The blocking rate wi e Static-max scheme is e highest among all due to its strictest admission criterion. Wi e proposed scheme, e blocking rate is between at wi e Static-max and Static-min schemes since e admission criteria are dynamically adjusted according to network loads (i.e., connection arrival rates). Figures 8, 9, and 10 show e achieved per-flow roughput of e rtps, ertps, and nrtps service classes, respectively, wi different CAC schemes. It is shown in ese figures at e Static-max and Static-min schemes provide steady roughput corresponding to e highestfidelity and minimum QoS, respectively. It is shown from Figures 7, 8, 9, and 10 at e Static-max and Static-min schemes cannot properly balance e tradeoff between network efficiency and user-perceived QoS due to using fixed admission criteria wiout e consideration of traffic loads. Consider e case of ertps traffic as shown in Figure 9 for example. Wi e Adapt scheme, e per-flow roughput keeps at e highest level when e network load is relatively light (l is less an 0.2); but, it degrades rapidly as e traffic loads get heavier (l is larger an 0.2) because e resources allocated to existing users are taken away to satisfy new connections QoS requirements. As e traffic load get much heavier such at l is larger an about 0.5, e achieved Connection blocking rate Adapt Static-max Static-min Connection arrival rate, λ Figure 7 The average connection blocking rates of e rtps, ertps, and nrtps traffic flows wi different CAC schemes.

14 Page 14 of rtps 100 Per-flow roughput (Kbps) Adapt Static-max Static-min Connection arrival rate, λ Figure 8 The achieved roughput of e rtps traffic class wi different CAC schemes. roughput can no more keep above e value for e lowest QoS (25 kbps). Wi e proposed scheme, e achieved roughput is varying from 38 to 66 kbps around wi e change of network loads, and remains at about 38 kbps when l is larger an 0.3. It is shown at e per-flow roughput can keep higher an e minimum QoS level. In Figures 8 and 10, we have similar observations regarding e rtps and nrtps traffic classes. The results demonstrate at e proposed dynamic CAC algorim can efficiently improve utilization of network resources and also assure adaptive QoS above e minimum requirements for admitted connections The new connection blocking rate and handoff dropping rate wi different BR schemes In is section, we conduct simulation scenarios which consider new connections and handoff connections to examine e performance of e proposed BR scheme wiout CAC. The compared BR schemes for performance evaluations are e FT and DT schemes [10] describedinsection3.2.thesimulationset-upinis scenario considers five classes of scheduling services, UGS, rtps, ertps, nrtps, and BE, each of which is associated wi specific maximum rates and minimum rates as shown in Table 3. To simply and fairly compare e performances of different BR schemes, our approach here considers e same fixed admission criteria as at used by e FT and DT schemes: e maximum rate is adopted as e admission criteria for UGS, rtps, and ertps, and e minimum rate is adopted for nrtps and BE [10]. Assume at e occurrences of e five scheduling services are wi uniform probabilities, wi which e arrival rates of new connections and handoff connections are l n and l h, respectively. We refer to [10] and adopt e following reshold setting. For e FT scheme, fixed /B is equal to 80%. In e DT scheme, min /B and max /B are set as 80 and 90%, respectively; dyn /B is set as 80% initially. In our scheme, min /B and max /B are equal to 0 and 90%, respectively, and consequently e initial value of ad /B is 45%. We examine e performance of different BR schemes wi different ratios of handoff arrival rates to new-connection arrival rates which can be 1:1 or 1:20 in particular. The performance metrics are indexed as e new connection blocking rates and handoff dropping rates. Figure 11a, b shows e new-connection blocking rate and handoff dropping rate, respectively, wi different

15 Page 15 of ertps Per-flow roughput (Kbps) Adapt Static-max Static-min Connection arrival rate, λ Figure 9 The achieved roughput of e ertps traffic class wi different CAC schemes. BR schemes as e ratio of handoff arrival rates l h to new-connection arrival rates l n is 1:1 (e total connection arrival rate l = l h + l n ). It is shown at e FT scheme provides e lowest handoff dropping probability and highest blocking rate when it adopts an FT for e BR. The DT scheme and our proposed scheme have similar performances in terms of bo blocking and dropping rates as ey determine e BR in an adaptive sense. Generally, e total amount of admitted connections wi e ree schemes is nearly e same. By comparison, our proposed scheme can admit a few more connections into networks an e FT and DT schemes as its increasing number of new connections is slightly larger an e decreasing amount of handoff connections. For example, when e arrival rate l is equal to 0.8 (l n = l h = 0.4) in particular, e total number of admitted new and handoff connections wi FT, DT, and e proposed scheme is 219, 218, and 225, respectively. In general wi various values of arrival rates l, our proposed scheme averagely can slightly increase e network efficiency by 0.75 and 1.47% in comparison wi e FT and DT schemes, respectively. Figure 12a, b shows e new-connection blocking rate and handoff dropping rate, respectively, when e ratio of handoff arrival rates l h to new-connection arrival rates l n is 1:20. The performance of e FT scheme is similar to at shown in Figure 11a, b at it has e lowest dropping and highest blocking rate among all. Wi regard to e DT scheme and our proposed scheme, it is shown at eir performances are quite dissimilar here, unlike what is shown in Figure 11a, b. We can observe at our proposed scheme provides a much lower blocking probability and a slightly higher handoff dropping rate an e FT and DT schemes (notice at e scales of e vertical axes in Figure 12a, b are different). Regarding is transmission scenario in which e occurrences of new connections are much more frequent an ose of handoff connections, e proposed scheme can significantly outperform e two schemes in terms of e amount of admitted connections. For example, when e arrival rate l is equal to 0.8 (l n = l h = 0.4) in particular, e blocking rate wi FT, DT, and e proposed scheme is about 0.54, 0.53, and 0.47, respectively, while e dropping rate wi e ree schemes is about 0, 0.007, and 0.009, respectively. Consequently, e total number of admitted new and handoff connections wi FT, DT, and e proposed scheme is 387, 399, and 442, respectively. The DT

16 Page 16 of nrtps Per-flow roughput (Kbps) Adapt Static-max Static-min Connection arrival rate, λ Figure 10 The achieved roughput of e nrtps traffic class wi different CAC schemes. scheme grants 12 connections more an e FT scheme, providing a slight increase of network efficiency by 3.20% over e FT scheme. It is shown at e DT scheme provides a raer marginal improvement over e FT scheme when e occurrences of handoff connections are relatively sparse. In is situation, e DT scheme may cause a waste of network resources as much as e FT scheme does. Wi e proposed scheme, e number of granted handoff connections is almost equal to at of e FT and DT schemes since a bandwid guard is used in our scheme to ensure a minimum BR for handoff users. On e oer hand, e New-connection blocking rate FT DT Handoff-connection dropping rate FT DT New-connection arrival rate, λn Handoff-connection arrival rate, λh (a) (b) Figure 11 As l h :l n is 1:1, (a) e blocking rates of new connections and (b) e dropping rates of handoff connections wi respect to different BR schemes. In comparison wi e FT and DT schemes, respectively, e proposed scheme slightly increases e network efficiency by 0.75 and 1.47%.