IEEE Globecom 2010 Workshop on Broadband Wireless Access

Joint Admission Control and Resource Allocation with GoS and QoS in LTE Uplink
Oscar Delgado
ECE, Concordia University Montreal, Qc, H3G 1M8, Canada Email: o delgad@encs.concordia.ca

Brigitte Jaumard
CIISE, Concordia University Montreal, Qc, H3G 1M8, Canada Email: bjaumard@ciise.concordia.ca

Abstract—In this paper, an admission control (AC) scheme is proposed for handling multiclass Grade of Service (GoS) and Quality of Service (QoS) in Uplink Long Term Evolution (LTE) systems. GoS requirement in conjunction with QoS has been seldom taken into account in previous admission control and resource allocation algorithms for LTE uplink. We propose a novel algorithm for handling the priorities while fulfilling the QoS objective of all granted requests. It corresponds to a solution that combines resource allocation and admission control properties to satisfy the GoS and QoS objectives. Call blocking probability, call outage probability, system capacity and number of effectively served requests are used as performance metrics. Numerical results show that it is possible to manage a priority scheme which satisfies the QoS constraints of all granted requests without any system capacity loss, when comparing to previous algorithms. Furthermore, the proposed AC algorithm gain, for the most sensitive traffic, can be around 20% over the reference AC algorithm. Index Terms—QoS, Priority, Admission Control, Scheduling, LTE, Uplink, Delay.

I. I NTRODUCTION Long Term Evolution systems (LTE) are designed to transfer all kinds of traffic, such as voice, video, data, etc, through packet switched networks based on Internet Protocol (IP). In order to provide high quality of service (QoS) in packet switched networks, Admission Control (AC) is needed. Its role is to accept or reject service requests in order to prevent network congestion, and to guarantee a certain level of QoS for on-going calls [1]. LTE systems aim to provide high data rates services. LTE standard is based on a decentralized architecture, with admission control functionality embedded in evolved Node-B at layer 3 [2], so it can utilize local cell load information to make the admission control decision. LTE systems also provide Quality of Service (QoS) guarantees to different types of services [3]. Nine Quality of service class identifiers (QCI) are defined, four Guaranteed Bit Rate (GBR) and five nonGBR. They are Conversational Voice, Conversational Video, Real-time Games, and Streaming Video for GBR, and IMS Signaling, Live Streaming, and TCP-based (e.g., www, e-mail, chat, ftp, p2p file sharing, progressive video, etc.) for nonGBR. Admission control techniques are not new in mobile networks [4], [5], [6]. Nevertheless, each new technology has its own characteristics, for example, in WCDMA systems, AC is

based on estimating and maintaining the intra-cell interference for granting a new request. As opposed to this, in LTE uplink intra-cell interference is not a main factor because of the use of Single Carrier Frequency Division Multiple Access (SC-FDMA) that could be considered as a special case of Orthogonal Frequency Division Multiple Access (OFDMA) [7]. Furthermore, in LTE uplink, requests are scheduled on a dynamically shared channel with fast adaptive modulation and coding (AMC) and hybrid ARQ (HARQ). Therefore, the AC algorithms for WCDMA system are not be suitable for LTE. An AC algorithm in downlink, which has similar PHY-MAC as uplink cannot be directly used in uplink because it has an additional degree of freedom i.e., user transmit power, which is different for all the users and can vary in each transmission time interval (TTI) due to power control. Admission control is important in order to maintain the QoS of on-going calls in a given cell, and to admit a new radio bearer only if all the existing calls and the new bearer can be guaranteed QoS. Hence, admission control has an important role to play for QoS provisioning. Nevertheless, admission control is tight to the scheduling scheme, so it may be advantageous to design a scheduling mechanism that uses the design criterion of the admission control. Many admission control schemes have been proposed with QoS constraints for multiclass services in wireless networks [8], [9]. But, as far as we know, very few have combined GoS and QoS concerns [10]. We propose here a novel admission control and resource allocation algorithm scheme that satisfies both GoS and QoS requirements. Our solution makes the GoS evaluation in the admission control and decides on the admission of a new request using not only the minimum throughput but also the maximum delay requirement. The resource allocation adapts dynamically to the number of requests in the system by assigning resources in a fair way. The proposed admission control algorithm has three advantages: 1) It is easy to implement as compared with other algorithms that include classes of service; 2) There is no specific threshold for different requests or traffic so we can distribute the traffic adaptively; 3) Most important of all, we can bring different classes of service according to their priority without reducing the spectrum efficiency and average cell throughput while fulfilling the QoS constraints of all granted requests.

978-1-4244-8865-0/10/$26.00 ©2010 IEEE

829

This paper is organized as follows. In Section II, we propose a novel admission control and a channel dependent scheduling algorithm that considers the priority and QoS requirements of different classes of traffic. Experiment results are presented and analyzed in Section III. Conclusions are drawn in the last section. II. A DMISSION C ONTROL AND S CHEDULING A LGORITHMS In this section, we propose a new admission control algorithm for LTE Uplink systems that not only satisfies the throughput and the delay requirements, but the traffic priority as well, when deciding on the admission of a new incoming request. A. Notations Let K be the set of granted requests, which is partitioned into traffic sets Ki such that K = Ki , where I is the index set of the different class of services, and Ki the set of requests min of class i ∈ I. Let Ri (resp. dmax ) be the minimum bit rate i (resp. maximum delay) required by class i of traffic. We denote by R the average overall throughput per TTI, assuming we maintain statistics over a given number of TTIs, distributed over the time period of concern. Indeed, instantaneous delays and throughputs are computed over a set of n successive TTIs and average values over a set of nw windows where each window is a set of n successive TTIs. B. AC Fair: A New Admission Control algorithm To grant a new incoming request, we need to guarantee min a minimum rate (Ri ) and a maximum delay (dmax ) for i each class i of traffic in addition to the blocking/dropping probabilities. Bit rates and delays are related as, if we increase the average bit rate, denoted by Ri , by assigning more blocks to some requests, then the average delay, denoted by di , also increases as we need to postpone the block assignment of the remaining requests. The proposed AC Fair algorithm checks if the current resource allocation can handle a new request while still satisfying the bit rate and delay requirements of all the active requests and of the new incoming request. Hence, the admission criterion for the new request is expressed as follows: di min min |Ki | Ri + Rinew ≤ R dmax i (1) i∈I We assume that each class i and type of request (new or handover) has a priority Qj , j = 1, 2, ..., pmax where index 1 corresponds to the highest priority and pmax to the lowest one. We next define the ratio Pj = Pbdj /PBDj where PBD is the target blocking/dropping probability, Pbd is the blocking ratio if it is a new request or the dropping ratio if it is a handover request measured in the time frame (of, e.g., nw windows of TTIs), i.e., number of rejected requests Pbd = (3) overall number of requests The AC Fair algorithm can be described as follows: Algorithm 1 AC Fair 1: Capture physical layer information and calculate average throughput and average delay of class i in the last n msec. 2: Set Ni (number of new granted requests of class i) to zero min 3: Set Rp to (di /dmax ) |Ki | Ri i i∈I 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15:

16: 17: 18: 19: 20:

Set Z (priority weight) to a value in [0, 1] Rank incoming requests according to Qj and select the requests in order of decreasing priority for each incoming request do min Collect traffic profile: Pj , Qj , Ri , dmax , ∀i, ∀j i max min or Ri ≥ Ri then if di ≥ di deny the request end if Find the smallest value of j∗ such that Pj ∗ > 0.9. If none, set j∗ to pmax if Rp < Z × R then j∗ = pmax end if if new incoming request k satisfies condition (1) and if it has priority Qj ≤ Qj ∗ or priority Pj > Pj−1 > Pj ∗ then Grant request k ; Ni ← Ni + 1 else Reject request k end if end for

i∈I

min min where Rinew = di /dmax Rinew is the minimum throughput i the system needs to provide to the new incoming request taking into account its class of service. It is dependent of the maximum delay that is guaranteed by the system, taking into account the set of already granted requests and the available bandwidth (i.e., available number of blocks or available bandwidth within each block per request). It is assumed that max di ≤ di

Parameter Z ∈ [0, 1] aims to prevent unnecessary blocking request due to previous congestion, for example if we define Z = 95%, and the system load is below that point, it will try to admit new request without checking the priorities. We assume that if the system is in a relaxed state all requests should have an equal chance of being admitted. C. RA: A New Scheduling Algorithm The packet scheduling is performed with a one step algorithm, called RA algorithm, that combines time-domain (TD) scheduling and frequency-domain (FD) scheduling. The RA algorithm selects the next request which will be multiplexed in time and frequency with the following metric. For a given request k of class i, it is defined as follows: dmax Rk (4) × imin fk = Ri dk

and

min R i ≥ Ri ,

(2)

as otherwise the new incoming request is denied.

830

where Rk is the average throughput of request k and dk is the average delay of request k. Indeed, the proposed RA algorithm allocates resource blocks with the aim of maximizing the throughput while making sure that requests will never experience a delay greater than the maximum allowed delay or a throughput smaller than the minimum throughput considering many classes of service. Algorithm 2 RA Algorithm 1: URB ← set of unassigned resource blocks delay R 2: Define F = {fk × δk × δk : k ∈ K} 3: Sort the values of F in ascending order. 4: for all k ∈ K do delay max 5: δk = 0 if Dk ≥ Di and k ∈ Ki , 1 otherwise R min 6: δk = 0 if Rk ≤ Ri , and k ∈ Ki , 1 otherwise 7: end for 8: x ← 1 9: while URB = ∅ do 10: Set k ⋆ to the index of the xth value of F 11: Assign RBx to request k ⋆ 12: Assign the channel such that max{Rk⋆ (RBc ) : RBc ∈ c URB } to request k ⋆ 13: URB ← URB \ {RBc} ; x ← x + 1 14: end while The reason for using the proportional fair scheduling metric fk instead of the typical proportional fair metric is that this new metric includes the delay requirement by giving more priority to requests that are closer to the maximum delay or to the minimum throughput. Assuming that the number of requests is always bigger than the number of RBs (as otherwise the block assignment is easy), the RA algorithm assigns RBs based on metric fk , using the fraction of the current throughput over the minimum throughput and of the maximum delay over the current delay. By doing so, we assign resource blocks to the requests with the most critical delay and throughput. III. P ERFORMANCE E VALUATION In this section, we present numerical results to study the performances of the proposed admission control and scheduling algorithms, we compare the proposed algorithm AC Fair with a tuned reference AC that also uses the new RA scheduling scheme. A. Simulation Model The simulation model consists of a single cell equipped with an omni directional antenna using SC-FDMA uplink based on 3GPP LTE system model. The throughput is averaged over 1000 TTIs, the duration of a TTI is 1msec. The total bandwidth considered is B = 5MHz, subdivided into 25 RBs of 12 sub carriers. All mobile users are assumed to transmit at the maximum power considered to be 125mW, the power is equally subdivided among all sub carriers allocated to the mobile in each TTI.

TABLE I PARAMETERS Parameter System bandwidth Pathloss Log-normal shadowing Shadowing correlation Penetration loss Fast fading TTI User maximum power Available MCSs Request arrival Request arrival rates Average call duration Simulation time Z Parameter Assumption 5 MHz (25 PRBs, 180 kHz per PRB) 128.1 + 37.6 log10(d in km) dB Standard deviation = 8 dB 1.0 for intra-site, 0.5 for inter-site 20 dB Typical Urban (TU3) 1 ms 125 mW M-QAM Poisson process [0.9, 0.96, 1.02, 1.08, 1.14, 1.2] calls/s 300 s 3600 s 95 %

TABLE II Q O S R EQUIREMENTS Services Voice Video Web QCI 1 2 7 Priority 2 4 6 Delay budget 100 ms 150 ms 300 ms Rate budget 64Kbps 256Kbps 16Kbps

For the throughput calculations, we consider that the data rate is upper bounded by the Shannon’s formula. The main simulation parameters are listed in Table I. Requests are generated in the system according to a Poisson arrival process. If the AC decision criterion proposed in Section II is fulfilled, the request is admitted, otherwise the request is blocked. To simplify the analysis without loss of generality, we consider three service classes (voice, video and web). Furthermore, the calls from the same service class can be either new in the cell or a handover from another cell, with handover calls having higher priority than new incoming calls as it is more annoying to drop a call than not having access to the network due to congestion; it might be also closer to the best conditions of the service level agreement (SLA). We choose arbitrarily the following priority scheme for our simulations handover voice > handover video > handover web > new voice > new video > new web, with handover voice as the highest priority and new web as the lowest priority. The traffic model settings are given in Table II Requirements of packet delays are considered as one fifth of the end to end packet delays in Table II. In simulations AC functionality is tightly coupled with scheduling algorithm as the QoS management is made in the scheduling part and priorities are handled by the AC algorithm. We consider three service classes, the traffic distribution is show in table III Since many admission control algorithm have been proposed for cellular systems, it is difficult to directly compare the performance of existing AC algorithms and the proposed
TABLE III T RAFFIC DISTRIBUTION Services Handover Voice, Video, Web Voice, Video, Web Distribution 8%, 4%, 8% 32%, 16%, 32%

831

AC algorithm. Here, we compare performance of our AC algorithm with a reference admission control one, due to [9], stated as follows:
K min min Rk + Rknew ≤ Rmax k=1

(5)

where Rmin is the minimum required throughput and Rmax is a predefined throughput that can be manually tuned to maximize efficiency. The reference AC of [9] decides to grant a new request if the sum of the Rmin of the new and of the active requests (users) is less than or equal to Rmax . B. Simulation Results The performance of the proposed AC and resource allocation algorithms is evaluated using blocking and outage probabilities as well as the average number of granted requests in the system, the average sum throughput in the cell and the delays. Blocking probability (Pb ) is defined as the ratio of the number of blocked request to the total number of request. Outage probability (Po ) is calculated as the ratio of the number of request not fulfilling their throughput and delay requirements to the total number of admitted request. Figure 1 (resp. 2) shows the blocking probabilities for our proposed AC algorithm (resp. the reference AC) as a function of the call arrival rate. We can draw the following conclusions. Firstly, the blocking probability increases as the call arrival rate increases. Secondly, the blocking probability is much higher for video calls as this class of traffic is more demanding in terms of bandwidth. Finally and most important in Figure 1 we notice that blocking probabilities follows the priority scheme required (ex. gold > silver > bronze), while in Figure 2 the priority distribution is random. This is because the blocking probability is not only dependent of the maximum capacity of the system but also of the throughput that the new requests are asking to the system.
Fig. 2. Blocking probability

the performance of any admission control algorithm we have to generate enough traffic to lead the system to the congestion region.

Fig. 3.

Outage probability

Figure 4 shows the average number of requests admitted in the system versus the arrival rate, we also consider the average number of request fulfilling the QoS requirements (useful number of request), for AC Fair the values of the average number of requests fulfilling the QoS requirements are slightly below of that of the reference AC algorithm, this is due to the fact that in the priority evaluation it can be loss some efficiency.

Fig. 1.

Blocking probability

Figure 3 depicts the outage probabilities versus the call arrival rate. We observe that the outage probabilities is best for the AC Fair algorithm than for the reference AC, this is due to the fact that AC Fair admit a new user only if its QoS can be satisfied, we must remember that in order to evaluate

Fig. 4.

Average number of granted requests

Figure 5 shows the useful sum throughput in the cell,

832

meaning the sum throughput of all the request fulfilling the QoS requirements. We observe that AC Fair algorithm has about the same behavior than the reference AC, values varies on a maximum of 2%. it is important to note that the maximum Rate Rmax of the reference AC algorithm is a manually tunable parameter and we are choosing the best value to make fair comparisons with our proposed AC algorithm, on the contrary AC Fair adapts dynamically to the traffic conditions.

Fig. 7.

Average voice delay

Fig. 5.

Average useful sum throughput

Figure 6 compares the useful sum throughput of video calls versus the arrival rate. We observe that the useful throughput of video calls for AC Fair algorithm is higher than the reference AC algorithm. This is because the outage probability of video calls is significantly higher, mostly due to the fact that video calls are more demanding (higher data rate).

so as to fulfill the QoS of the new and existing requests, and the Resource Allocation assigns RB’s in a fair way such that the throughput and delay is adaptively adjusted according to the traffic load. System performance is evaluated using simulations. Results show that although the total sum throughput is not improved, the proposed AC algorithm gain for the most sensitive traffic can be around 20% over the reference AC algorithm without sacrificing the overall system capacity and at the same time guarantying GoS and maintaining the basic QoS requirements for all the admitted request. R EFERENCES
[1] Requirements for Evolved UTRA (E-UTRA) and Evolved UTRAN (EUTRAN), 3GPP Technical Specification Group Radio Access Network, December 2008, 3GPP TR 25.913 v.8.0.0. [2] H. H. and T. A., LTE for UMTS - OFDMA and SC-FDMA Based Radio Access. Wiley, June 2009. [3] Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved Universal Terrestrial Radio Access Network (E-UTRAN),Overall description, 3GPP Technical Specification Group Radio Access Network, December 2008, 3GPP TS 36.300 v.8.7.0. [4] M. Ahmed, “Call admission control in wireless networks: a comprehensive survey,” Communications Surveys Tutorials, IEEE, vol. 7, no. 1, pp. 49 – 68, qtr. 2005. [5] H. Lei, M. Yu, A. Zhao, Y. Chang, and D. Yang, “Adaptive Connection Admission Control Algorithm for LTE Systems,” May 2008, pp. 2336 –2340. [6] S. P. R. Babu H.S., G. Shankar, “Call admission control approaches in beyond 3g networks using multi criteria decision making,” IEEE wireless communications, cicsyn, First International Conference on Computational Intelligence, Communication Systems and Networks, pp. 492–496, September 2009. [7] M. Rumney, “3GPP LTE: Introducing single-carrier FDMA,” Agilent Technology, Tech. Rep., January 2008. [8] S. J. Bae, B. Choi, M. Y. Chung, J. J. Lee, and S. Kwon, “Delay-aware call admission control algorithm in 3GPP LTE system,” TENCON, IEEE Region 10 Conference, pp. 1–6, November 2009. [9] F. D. C. K. I. P. P. E. M. M. Anas, C. Rosa, “Combined Admission Control and Scheduling for QoS Differentiation in LTE Uplink,” IEEE 68th Vehicular Technology Conference (VTC), pp. 1–5, September 2008. [10] P. Hosein, “A Class-Based Admission Control algorithm for shared wireless channels supporting QoS services,” in Fifth IFIP-TC6 Conference on Mobile and Wireless Communication Networks, 2003, pp. 81 – 85.

Fig. 6.

Average video useful sum throughput

Last, Figure 7 shows the average delay of each class versus the total number of requests in the system. We observe that MC-SA and SC-SA meet the maximum allowed delay. One important thing we can point out is that in order to meet the video requirements, both systems assign resources to video calls at every TTI, nevertheless MC-SA is the only one who guarantee the minimum throughput. IV. C ONCLUSION In this paper, we developed a combined Admission Control and Resource Allocation algorithm for handling multiclass GoS and QoS in LTE Uplink systems. The AC determines if a new request can be accepted based on its priority and in the real minimum throughput and delay the system can offer

833