Wireless Systems. Shengming ... present (i.e., the center of shadow) and all its adjacent cells. ... For exam- ple, a mobile-phone user sitting in a co ee bar may.
An Enhanced Distributed Call Admission Control for Wireless Systems Shengming Jiang Centre for Wireless Communications National University of Singapore
Danny H.K. Tsanga , Bo Lib aDept. of Electrical and Electronic Engineering b Computer Science Department Hong Kong University of Science and Technology
Abstract
Call admission control (CAC) is one of the key elements to provide quality of service (QoS) guarantee in wireless mobile networks. A good CAC scheme should eciently support hando s and maintain a high utilization of the scarce radio resource while it should also be simple for implementation without loss of applicability. Since the traditional guard channel policy is not optimal for a hard constrained hando dropping probability, there are some new proposals in the literature. Among them, the distributed call admission control seems to be simpler and more ecient with regards to the above requirements. However, there are some problems in the original scheme which make the implementation complex and lead to performance degradation. In this paper, we will rst discuss simpli cations to the original scheme, and then further enhance the modi ed scheme by considering the di erence of mobility support requirement between high and low mobility users. The simulation results shows that the enhanced scheme gives better performances.
Keywords: Call admission control, Mobility classi cation and QoS guarantee.
1 Introduction Future mobile communication systems will permit a user to communicate from any where at any time with on-demand quality of service (QoS) for di erent applications such as data, voice and video. One of the key elements to provide QoS guarantee is call admission control (CAC). CAC denotes the process to make a decision for new admission according to the amount of available resource versus user's QoS requirements and the e ect upon the QoS of the existing calls imposed by new admission. Di erent from the CAC for the wireline multimedia networks, the CAC for wireless systems is This work is supported by Hongkong Telecom Institute of Information Technology under the HKTIIT 93/94.EG01 grant.
more complex and critical due to mobility and scarce radio resource. Therefore, the CAC for wireless systems has to eciently support hando and maintain a high resource utilization. Furthermore, a good CAC scheme needs also to be simple for easy implementation without loss of applicability. The traditional guard channel policy [8] and the fractional guard channel policy [7] determine a long-term number of guard channels for hando s by just considering the status of the local cell. And assumptions are needed for call arrival processes, call lifetime and channel holding time to calculate the number of guard channels. Recently, there are some other CAC schemes such as the shadow cluster concept [3] and the distributed call admission control (D-CAC) [2]. These two schemes reserve resource dynamically for hando s and require less assumptions. The idea of the shadow cluster concept is that \every mobile with an active wireless connection establishes an in uence upon the cells in the vicinity of its current location and its direction of the travel" [3]. The coverage of a shadow for a given active mobile consists of the cell where the mobile is currently present (i.e., the center of shadow) and all its adjacent cells. This area moves as the mobile hando s to other cells. The eciency of this scheme heavily depends on the accuracy of algorithms to predicate the potential movement of mobile hosts. Another problem is the cost given by this scheme because the operation of the shadow cluster is needed for every new call request and hando . D-CAC can be regarded as a dynamic guard channel policy subject to a hard constrained CDP. Compared to the above schemes, D-CAC is simpler and needs less assumption. However, call admission process with DCAC needs to be Poisson to calculate call admission threshold. This assumption seems to be unpractical for implementation and con icts with the pre-calculated call admission threshold. Another problem is how to set the CAC period T to meet the originally assump-
tion that T should be short enough so that the probability that a call hands o more than once during this period is negligible. Therefore, in Section 2, following an introduction to the D-CAC in Section 2.1, we will rst discuss simpli cations to the original D-CAC by removing the Poisson assumption mentioned above and describe a measurement-based method to determine T. On the other hand, the D-CAC makes resource reservation for a new admitted call in both local and adjacent cells. As discussed in [1, 4, 5, 6], a mobile user may have di erent mobility support requirements in di erent communication environments. For example, a mobile-phone user sitting in a co ee bar may never need the inter-cell hando support during communication while the hando support is indispensable for a taxi passenger using a mobile-phone. Obviously, for requests from \sitting-users", resource allocation in local cell is enough to guarantee QoS without extra resource allocation in other cells. Therefore, in Section 3, we will further discuss an enhancement to the simpli ed D-CAC by considering di erent requirements on mobility support. Finally, we conclude the paper in Section 4.
Call Terminations
Cl0
Cl
Cn
Cr
Cr0
l0
l
n
r
r0
Handoffs
Handoffs
Call Arrivals
Figure 1: One-dimensional cell array for D-CAC arise from this assumption. i) How to determine mean admission rate for new calls () with a pre-calculated threshold? ii) How to implement a Poisson admission process with a pre-calculated threshold? For implementation to satisfy this assumption, the interval between two consecutive new admissions occurring in a cell needs to obey the exponential distribution. Suppose a new call have been admitted at time t1 . At the same time, there are empty channels and a non-zero admission threshold that is valid until t in the cell. The next admission should occur at t2 = t1 + , where is generated according to the exponential distribution. Consider t2 t . Since the new call arrival process is also stochastic, a new call may arrive i) before t2 , ii) at t2 or iii) after t2 . Case `ii' is ideal but rare. In case `i', the new call will be queued until t2 for admission with compliance to the above Poisson-assumption. This will cause resource wastage. However, it is complex to deal with case `iii'. In this case, a new t2, t2, is needed by regeneration of at t2 . However, t2 may be still smaller than the new call arrival time. Therefore, the above process may need to be iterated until either case `i' or case `ii' occurs. On the other hand, a devise is needed to determine T since the T setting heavily a ects the accuracy of the calculation for call admission threshold and will further a ect QoS guarantee as shown by the simulation results in [2]. 0
0
2 Simpli ed D-CAC
2.1 Introduction to D-CAC
D-CAC [2] determines the admission threshold for a radio cell to admit new calls during a CAC period T by considering the number of active calls in both the local and the adjacent cells. The CAC operation performs every T, and T is supposed be short enough so that the probability that a call hands o more than once during T is negligible (called T-assumption henceforth). Consider one dimensional cellular array consisting of ve cells denoted by Cl0 , Cl , Cn , Cr and Cr0 from the left to the right as depicted in Figure 1. The conditions for cell Cn to admit a new call during T are as follow: i) The new admission in cell Cn cannot a ect the CDP guarantee promised to the hando calls coming from cells Cr and Cl to cell Cn during T; ii) Since the new calls admitted by cell Cn may hando to cells Cr or Cl , the new admission in cell Cn cannot a ect the CDP guarantee for the hando s going to cells Cr and Cl from Cn during T. Please refer to [2] for more details on D-CAC. D-CAC needs no assumption on hando and new call arrival processes while the admission process is supposed be Poisson to calculate the call admission threshold for a CAC period T. This assumption seems to be unpractical for implementation and con icts with the pre-calculated admission threshold. Two questions
0
0
2.2 Simpli cations to D-CAC To resolve the above problems in the original scheme, we need to remove T from the original equations in [2]. T comes from the estimation of the number of new admitted calls during a CAC period T. Consider cell Cn in Figure 1 calculating call admission threshold at t0 for the next CAC period T (Ntn ). Denote by l0 , l, n, r and r0 the numbers of calls existing in cells Cl0 , Cl , Cn, Cr and Cr0 at time t0, respectively. The threshold is calculated through estimation on the distribution of the number of calls to be present in cells Cn (Nn ), Cr (Nr ) and Cl (Nl ) at t0 + T. According to admission condition `i' mentioned earlier, a threshold n1 imposed 2
by Cn can be calculated. Another two thresholds n2 and n3 imposed by Cr and Cl , respectively, can be also obtained according to admission condition `ii'. Then, Ntn = minfn1 ; n2; n3g. In fact, Nr consists of the numbers of calls among r calls to remain in Cr at t0 + T, the new admitted calls and the hando s coming from Cr and Cl between t0 and t0 + T. Only the former two types of calls will generate hando s during a CAC period due to the T-assumption. The sum of those two numbers is bounded by R = max(r; Ntr ), where Ntr is the call admission threshold for Cr . Therefore, we suggest to use R instead of the sum to avoid the estimation of the new admitted calls so as to remove T used in the original equations. Similarly for Cl , L = max(l; Ntl ) is also adopted, where Ntl is the call admission threshold for Cl . Di erent from the original D-CAC, we think it is also necessary to consider the hando s introduced by the new calls admitted by Cr and Cl in calculation of n1 . Therefore, r and l in the equations in [2] are also replaced by R and L, respectively. Then, we have the following formulas to calculate n1, n2 and n31 n1 = 2p1 [a2 (1 ? ps ) + 2N ? pm (R + L) ? ps (1) a a2(1 ? ps)2 + 4N(1 ? ps ) ? A] n2 = 2p1 [a2(2 ? pm ) ? 2R0pm ? 4(Rps ? N) ? m p a a2(2 ? p
T-assumption, the shorter T, the more accurate the calculation is for the call admission threshold. This will be re ected by the approach between the target CDP and the measured CDP. Therefore, set the initial T small, then increase T until the measured CDP is close to but not bigger than the target CDP. However, a very long T setting can also give a small CDP that meets the target CDP since a long T cannot increase the admission threshold in a cell immediately if there are many calls have left the cell. This will give a low CDP but a high CBP. This phenomenon can be observed by the system because there will be a lot of unused empty channels while new call requests are still dropped. In this case, T should be decreased until another T setting gives a CDP smaller than but close to the target CDP, and also gives a smaller CBP. We also propose a measurement method for pm and ps instead of the mathematical calculation used in [2] since the calculation requires assumptions on distributions for call lifetime and channel holding time. The measurement can be operated periodically and the nal result is the weighted accumulation of all the periodical results. It is easy for a cell to observe the events of hando s and termination of calls. Therefore, a cell can count the calls that hando to other cells (Kho ), the calls that terminate in the local cell (Kte ) and total calls present in the local cell (K) during a measurement period Tm . The probabilities that a call handso to other cells (PTmm ) or terminates (PTtm ) during a measurement period can be approximated by PTmm = Kho =K and PTtm = Kte =K. The probability that a call remains in the local cell during a measurement period, PTsm , is given by 1 ? PTmm ? PTtm . The weighted accumulation of PTmm and PTtm from 0 to t+Tm can be generally represented by Mt+Tm = Mt +(1 ? )MTm , where Mt is the weighted accumulation from 0 to t, MTm is the measured result at t + Tm , and a is a weight between 0 and 1.
m )2 + 16N + f(R)]
(2) 1 n3 = 2p [a2(2 ? pm ) ? 2L0 pm ? 4(Lps ? N) ? m p (3) a a2(2 ? pm )2 + 16N + f(L)]
where, A = pm (pm ? 2ps )(R+L), f(x) = 8pm (xps ? N) ? 16xp2s , R0 and L0 , the maximums between the number of the active calls and the admission thresholds respectively for cells Cr0 and Cl0 , are used to replace E(n) (the mean number of active calls) used in [2]; ps and pm denote the probability of a call to remain in the same cell and the probability of a call handing-o to other cells, respectively.
2.4 Simulation results This subsection investigates reasonable T settings for the target CDP guarantee. T is set manually, and pm and ps are obtained by measurement. Tm is set to a CAC period and = 0:5 for measurement of pm and ps in all the simulations used in this paper. The simulation model used here is the same as the one used in [2], i.e., 10 one-dimensional cells arranged on a circle. The system capacity is set to 50 channels in each cell and the target CDP is 0.01. The new call arrival process is Poisson with mean arrival rate = 0:4=s. Both the call life time and channel holding time are set to be exponentially distributed.
2.3 Measurement for T , pm and ps
An important issue is how to determine T to guarantee CDP while to keep resource utilization high. It is dicult to calculate T mathematically without other assumptions. A practical way to estimate T is to devise an algorithm to adjust T automatically according to the measured CDP and CBP. According to the 1 Equations 1, 2 and 3 were derived the same way as used in [2] with consideration of the substitutions mentioned above. However, the detail derivation is ignored here due to the volume limitation.
3
Figures 34 show CDP, CBP and channel utilization () for di erent T-settings with mean call life time = 500s. As shown in Figure 3, the common T-range to guarantee the target CDP 0:01 for the 3 settings of mean channel holding time is about between 0 and 40s. In case of T > 40s, the values of CDPs in this range become irregular and most of them go beyond the target CDP. The reason is a long T-setting will invalidate the T-assumption and enlarge the inaccuracy in the calculation for the call admission threshold. Still some CDPs are below the target CDP. This may be due to a long T-setting hesitates the system to increase the admission threshold when there are many calls have left the system. This gives more chance for hando s to be accepted but a higher CBP and a lower as shown in Figures 3 and 4. 0.95
0.94
0.9
Channel utilization
0.88
Call Blocking Probability
40 50 60 CAC Period T (second)
0
o: h = 100s X: h = 50s
0.5
10
20
30
40 50 60 CAC Period T (second)
70
80
90
100
Figure 2: CBP versus T and mean channel holding time
h: Mean Channel Holding Time
0.18 0.16
Call Dropping Probability
30
+: h = 150s o: h = 100s X: h = 50s
0.1 0.08 0.06 0.04 0.02
10
20
30
40 50 60 CAC Period T (second)
70
80
90
70
80
90
100
In the following, we will discuss an enhancement to the simpli ed D-CAC by considering the di erence on mobility support existing between low and high mobility calls. Denoted by 1 and 2 the mean call arrival rates for low and high mobility calls present in a given cell, respectively, and = 1 + 2 , where indicates the total mean new arrival rate. Suppose the entire capacity of a cell, N, is statically divided into N l and N h for low and high mobility calls, respectively, with N l =N h = 1 =2 and N = N l + N h . Then, N l = 1+12 N and N h = 1+22 N. For low mobility call requests, resource allocation is needed only in the local cell where the user submits the request. The admission threshold for low mobility calls (Ntl ) is N l , which, however, can be shared by hando s. N h channels are shared by high mobility calls and hando s generated by the admitted high mobility calls. The simpli ed D-CAC is used to set the admission threshold for the high mobility call. To this end, N, R, L, R0 and L0 in Eqs (1) (3) are replaced with N h , Rh , Lh , Rh0 and Lh0 , respectively. Rh , Lh , Rh0 and Lh0 are the maximums between the number of active high mobility calls and the admission thresholds for high mobility calls for cells Cr , Cl , Cr0 and Cl0 , respectively. Denoted by Noh the number of occupied channels for high mobility calls, Nol the number of occupied channels for low mobility calls and and Nhol the number of low mobility channels occupied by hando s in a given cell. The simulation model used here is the same one as used in Section 2.4. Figures 5 and 6 compare performances given by the scheme with mobility classi cation (i.e., the enhanced one) and the one without mobility classi cation (i.e., the modi ed D-CAC mentioned earlier). Two scenarios with di erent portions between low and high mobility calls are under consideration as follows: a) 1 : 2 = 0:25 : 0:75 and b)
h: Mean Channel Holding Time
0.55
0 0
20
3 Enhanced simpli ed D-CAC
+: h = 150s
0.12
10
Figure 4: versus T and mean channel holding time
0.7
0.14
0.82
0.74 0
0.8
0.2
0.84
0.76
0.75
0.45 0
0.86
0.8
0.9
0.6
+: h = 150s o: h = 100s X: h = 50s
0.78
0.85
0.65
h = Mean Channel Holding Time
0.92
100
Figure 3: CDP versus T and mean channel holding time 4
0
0.014
MC = Mobility Classification 0.9
0.012
0.008
*: with MC
λ1:λ2=0.25:0.75
o: without MC
λ1:λ2=0.25:0.75
+: with MC
λ1:λ2=0.75:0.25
X: without MC
λ1:λ2=0.75:0.25
0.85
0.8 Channel utilization
Call Dropping Probability
0.01
0.006
0.004
0.75
0.7
0.65
0.002
λ1:λ2=0.25:0.75
*: with MC o: without MC
λ1:λ2=0.25:0.75
+: with MC
λ1:λ2=0.75:0.25
X: without MC
λ1:λ2=0.75:0.25
0.6 MC = Mobility Classification
0 0
5
10
15 20 25 CAC Period T (second)
30
35
0.55 0
40
5
10
15 20 25 CAC Period T (second)
30
35
40
Figure 5: CDP versus mobility classi cation
Figure 6: versus mobility classi cation
1 : 2 = 0:75 : 0:25. The common T-range for both schemes to guarantee the target CDP 0.01 is between 0 and 29s as shown in Figure 5. However, the CDPs given by the scheme with mobility classi cation are further below the target CDP 0:01 for both scenarios while the CDPs given by the scheme without mobility classi cation in scenario `a' goes beyond 0:01 quickly when T > 30s. Hence, the T used by the scheme with mobility classi cation can be adjusted longer to reduce the CAC operation while it can still guarantee the target CDP. The more signi cant improvement given by the enhanced D-CAC is the high channel utilization, and this improvement will increase with increase of the portion of low mobility calls in the total call arrival rate as shown in 6. Although the CDP given by the scheme without mobility classi cation is also smaller than the target CDP for scenario `b' in case of T > 30s, the channel utilization given by this scheme is much lower than that given by the scheme with mobility classi cation for the same scenario as shown in 6. The reason is that the some of the resource reserved for hando s by the scheme without mobility classi cation is wasted since low mobility calls does not hando as mentioned earlier. This wastage becomes heavy as the portion of low mobility calls in the total call arrival rate is high (e.g., case `b') as shown in the gure.
both parts were investigated by simulation. We have also proposed algorithms to nd the reasonable T and measure pm and ps . Therefore, the enhanced D-CAC can be used in general environments. The further research will extend the enhanced scheme for multimedia environments, where applications will require di erent amounts of bandwidth with di erent QoS requirements.
References [1] S. M. Jiang, Danny H.K. Tsang and Bo Li, SubscriberAssisted Hando Support in Multimedia PCS, ACM Mobile Computing and Communication Review, September 1997, Vol.1 No. 3, pp29-36. [2] M. Nagshineh and M. Schwartz, Distributed Call Admission Control in Mobile/Wireless Networks, IEEE JSAC, Vol. 14, No. 4, May 1996, pp711-716. [3] D. A. Levine et al., A Resource Estimation and Call Admission Algorithm for Wireless Multimedia Networks Using The Shadow Cluster Concept, IEEE/ACM Transaction on Networking, Vol. 5, No.1, February 1997, pp1-12. [4] Victor O.K. Li and X. M. Qiu, Personal Communication Systems (PCS), Proceedingsof the IEEE, vol. 83, No. 9, September 1995, pp1210-1243 [5] D. Hong, S.S. Rappaport, Priority oriented channel access for cellular systems serving vehicular and portable radio telephones, IEE Proceedings, Vol. 136, Pt. I, No. 5, October 1989, pp339-346. [6] S.S. Rappaport, Blocking, hand-o and trac performance for cellular communication systems with mixed platforms, IEE Proceedings, Vol. 140, No. 5, October 1993, pp389-401. [7] R. Ramjee, R. Nagarajan and Don Towsley, On Optimal Call Admission Control in Cellular Networks, INFOCOM'96, pp43-50. [8] E. C. Posner and R. Guerin, Trac Policies in Cellular Radio that minimize blocking of Hando calls, ITC-11, Kyoto, 1985.
4 Conclusion The main contributions of this paper consists of two parts. The rst one is simpli cations to the original D-CAC. The modi ed scheme is simpler than the original one for ease implementation. The second one is to merge the mobility classi cation approach into the modi ed D-CAC. The enhanced scheme is more ef cient to use the scarce radio resource with guarantee of call dropping probability. The performances of 5
