See it at CiteSeerX

1 downloads 83 Views 136KB Size Report
C11. Fig. 1. Example OVSF code tree. We consider that only calls with rates equal to the OVSF code .... CP, HP3 and HP3R, especially as the load increases.
Comparison of Code Reservation Schemes at the Forward Link in WCDMA Angelos N. Rouskas1 and Dimitrios N. Skoutas2 Department of Information and Communication Systems Engineering, University of the Aegean, Samos 83200, Greece {1 arouskas,2 d.skoutas}@aegean.gr

Abstract - We examine resource reservation schemes for the management of orthogonal variable spreading factor (OVSF) codes at the forward link of 3G mobile communications systems employing WCDMA. Like in every multiservice network, different rate calls in 3G systems will perceive dissimilar system performance if no measures are taken and the channelization code tree is treated as a common pool of resources. The paper introduces complete sharing, complete partitioning and hybrid partitioning strategies to manage the code tree, and studies the performance in terms of blocking probability per traffic rate class and utilization of codes. It turns out that fair access to codes by different rate calls and code utilization are conflicting goals, and that hybrid schemes can provide a compromise between these two extremes. Keywords : Third generation; WCDMA; OVSF; code reservation. 1. I NTRODUCTION

High data rate and variable data rate services ranging from 144Kbps to 384kbps for wide and up to 2Mbps for low coverage and mobility are anticipated in 3G mobile communication systems. WCDMA supports multiple rate transmission using single orthogonal channelization codes with variable spreading factors (SF), the so-called OVSF codes. Since OVSF codes are valuable and scarce resources, OVSF code assignment at the forward link of the WCDMA wireless interface has received great attention recently. In [1], the situation of code blocking is introduced, where an incoming call may be blocked even if there is spare capacity that can handle the new call. This phenomenon is encountered only on incoming calls with rate greater than the lower code rate available and is due to the scattering of free codes along the tree. The higher the rate of the incoming call the worse the code blocking, in other words higher rate calls are facing severe limitations on accessing the codes at increasing loads. Code reassignments at the arrival of new calls can be used as a means to completely eliminate the problem of code blocking, and the paper focuses on procedures that perform the minimum possible number of reassignments. The problem of selecting the best code, among sev-

0-7803-7605-6/02/$17.00 ©2002 IEEE.

eral available candidates, at the arrival of a new call or during the reassignment phase, is studied in [2], [3]. This is another way to prevent (or at least alleviate) the appearance of code blocking, by allocating those codes that will cause the least segmentation of the tree capacity. Code blocking, however, is only part of the problem encountered on the assignment process, where higher rate calls are experiencing severe blocking compared to the lower rate calls, especially at medium to high level loads. This situation is common in multirate service networks and is studied extensively in the area of ATM networks, fixed or wireless (see for example [4] and the references therein). To cope with this problem, priority based schemes for different service classes are presented in [5], while a new protocol that dynamically assigns OVSF codes on a timeslot basis to provide for data rate guarantee for bursty traffic is presented in [6]. In this paper, we introduce resource reservation strategies to provide for fair access of different rate calls to the available limited set of OVSF codes at the forward link of WCDMA systems. The reservation is made according to some prior knowledge of network operators about the traffic distribution of different classes. At the two extremes, we have (a) the complete sharing (CS) strategy, in which the whole code tree is considered as a common pool of resources, and (b) the complete partitioning (CP) strategy, where each traffic rate class is allocated a separate partition of the code tree for use. Between the two extremes hybrid partitioning (HP) strategies are developed and evaluated. The performance of resource reservation strategies are studied in terms of blocking per class and code utilization. Simulation results show that (a) fair access to codes, in other words similar blocking probabilities, is assured more by CP and less by HPs, (b) HPs are more immune against traffic variations than CP, which is usually optimized for certain traffic mixes, and (c) CS with or without reassignments, provide slightly higher code utilization than CP at the expense of very dissimilar performance experienced by different rate calls. The paper is organized as follows. The system model is pre-

sented in Section 2. The long term resource allocation process, as well as the real time assignment operation are developed in Section 3. Finally, Section 4 presents and discusses the simulation results. 2. S YSTEM M ODEL

The OVSF code tree is a binary tree with h layers, where each node represents a channelization code Ckl , k = 1, . . . , h, l = 1, . . . , 2h−k [7], [8]. The lowest layer is the leaf layer and the highest layer is the root layer. Codes of the same layer, as well as codes that do not lie in the same path from a leaf to the root of the tree, are orthogonal. The lower the layer of a code, the higher the spreading factor and the lower the code rate. If R denotes the lower (leaf) code rate and SFmax denotes the maximum spreading factor of the leaf layer codes, then the rate of the root layer code is 2h−1 R, and its spreading factor is SFmax /2h−1 . Figure 1 depicts a 5-layer code tree. Due to the orthogonality constraint, the assignment of a code is possible, if and only if none of its ancestor codes and none of its descendant codes are occupied. Once a code is assigned, all of its ancestors, as well as all of its descendants are blocked and can be used after the code is released. For example, the assignment of code C34 immediately blocks its ancestor codes C51 , C42 , and its descendant codes C27 , C28 , C1,13 , C1,14 , C1,15 , and C1,16 , as shown in Figure 1. The code tree capacity is 16R, the used capacity is 8R, and the remaining capacity is 8R.

C41 C31

C51 - SFmax/16 C42 - SFmax/8 C32 C33 C34 - SFmax/4

C21 C11 occupied

blocked

C28 - SFmax/2 C1,16 - SFmax free

Fig. 1. Example OVSF code tree

We consider that only calls with rates equal to the OVSF code rates are introduced to the system, and that only single OVSF code operation is possible. The arrivals of different rate calls are modeled as independent Poisson processes, and the call duration is exponential. Calls with rate requirement iR, where i = 2k−1 , are arriving at a rate λi and their call duration is exponentially distributed with mean 1/µi . The offered load of rate class iR is then αi = λi /µi . 3. A SSIGNMENT P ROCESS

We identify two levels of code management. • At the resource reservation level, the network operator is given the possibility to reserve certain parts of the OVSF

code tree for certain rate calls. At the call level incoming calls are assigned to free codes of the part that those calls have access to. In other words, the former is a resource planning, while the latter is a real-time operation. •

3.1. Allocation Strategies at the Resource Reservation Level Complete Sharing - (CS) : When there is no partition of the code tree and each incoming call can have access to every code (at the corresponding rate level of the tree) that preserves the orthogonality constraint, then we have the complete sharing strategy. This strategy is the only one studied so far in the context of OVSF code assignment at the forward link of WCDMA interface. Complete Partitioning - (CP) : When different rate calls have access to different non-overlapping partitions of the code tree, then we have the complete partitioning strategy. In this scheme, there are some issues that need to be further clarified: 1) The problem of how many codes each partition must acquire: Uniform partitioning is expected to work well under uniform traffic loads, but it will under utilize codes at different traffic mixes. Consequently, partitioning must take into account some prior knowledge about the expected offered traffic distribution. 2) Even if such knowledge is exploited, a non-uniform partition may work well for some offered traffic intensities but not for others. Thus, a decision has to be taken about the traffic level the partition should be optimized for. 3) Due to the binary code tree structure, the coarse rate values (powers of two times the basic rate), and the fixed size of the tree, the partitioning may be also quite coarse. To clarify this, let us denote by ni the number of rate iR codes of the corresponding partition, i = 2k−1 . Then, in a tree with P 256 leaf codes, the following constraint must hold: i i × ni = 256. Fortunately, since the size of the tree is small, the number of possible combinations of ni ’s is limited and a sequential search to find the most suitable is possible. The problem of selecting the most suitable distribution of ni ’s can be described as: Given the offered traffic ai of each class, and the size of the code tree, find the combination of ni ’s such that the blocking probabilities experienced by each traffic class is as similar as possible. The blocking probability for each trafic class can be obtained by the Erlang B(ni , ai ) formula. One straightforward way to identify the optimum combination of ni ’s is to sequentially search among the set of all possible combinations and choose the one that yields the minimum standard deviation of the corresponding blocking probabilities. Any such partition should pre-

serve the fairness property ni > 0. Hybrid Partitioning - (HP) : Between the two extreme strategies, hybrid partitioning schemes can be devised. For example, one solution might be the division of the code tree in two parts. In the first part, CP may be applied, while the second part is left for CS. However, due to the abrupt difference on the size of code rates, this type of hybrid strategies are not expected to behave well, because of the aggressiveness of the lower rate calls. In other words, at medium to high loads, the lower rate calls will flood the common pool, and higher rate calls will be soon confined to their own partitions. Thus, from a CP scheme we introduce the following hybrid scheme, which favors higher rate over lower rate calls: The calls of rate iR can have access not only on the partition of their own but also on the partition of the immediately lower rate iR/2 calls. This HP strategy is denoted by HP1. Similarly, if calls of rate iR can have access on the partition of iR/2 and iR/4 calls the HP strategy is denoted by HP2, and so on. Within a shared area, the codes are treated as a common resource by the calls that have access to that area, just like in the CS strategy. However, incoming calls of rate iR first search for empty codes in their partition and if no free codes are found, the search is continued to the common area of iR and iR/2 calls. The aforementioned CP and HP strategies are illustrated in Figure 2. Complete Partitioning (CP)

1R codes area

2R codes area

4R codes area

8R codes area

Hybrid Partitioning 1 (HP1)

1R codes area 2R codes area

4R codes area

8R codes area

Hybrid Partitioning 2 (HP2)

1R codes area 2R codes area

4R codes area

8R codes area

Fig. 2. Complete partitioning vs hybrid partitioning 1 and 2

3.2. Code Selection Schemes at the Call Level During system operation, and because of the statistical nature of the arrival and departure processes, the occupied codes will be randomly scattered across the code tree (CS case) or within the common area of codes (HP cases), if no countermeasures are taken and the real-time allocation of codes is performed either randomly or even in an ordered fashion. The obvious result will be fragmented tree or common areas capacity and limited appearances of higher free codes. One way to alleviate this phenomenon is by applying clever code selection schemes at the arrival of a new call, when more than one codes of the required rate are available. We can use any of the crowded-first or weighted code selection schemes proposed in [2], [3]. They both yield nearly identical results in terms of code blocking and code utilization. According to the second scheme, which is used in the simulations below, the selected code is the one which blocks the least possible higher free codes. For example, in the case of an incoming call of rate 2R in the tree of Figure 1, the candidate codes to receive the new call are C21 , C22 , C24 , and C26 . The weighted selection scheme ignores C21 , and C22 , and selects either C24 , or C26 , depending on the ordered criterion used, leftmost or rightmost. If such a scheme is applied, after each code allocation the largest available code will be the largest possible. In other words, the scheme will not block higher layer codes, unless there is no alternative, thus alleviating the problem of scattered codes. Obviously, such a scheme is required in the CS and HP resource reservation strategies only, since in the case of CP, each traffic class has its own pool of codes to select from and there is no scattering of codes within each pool. 3.3. Code Reselection Schemes at the Call Level Reassignment schemes are necessary when the capacity of the tree is enough to carry the incoming call, but no code of the required rate is available. In this case, a code of the required rate can become available, by transfering all the ongoing calls at the subbranch of the code to some other branches of the tree. Since the reassignment of a code may involve further reassignments of codes in other branches, the only way to ensure that the minimum number of reassignments be performed is to exhaustively search for the best candidate code with a recursive procedure. However, this is not possible during call setup time and a heuristic algorithm is required. The idea behind the code pattern search, mentioned in [1], is to select the code which has the least number of occupied codes in its subbranch, and if more than one codes are still candidates, then the one with the least occupied capacity in its subbranch is selected. Reassignments of lower rate codes are treated like new calls and may induce further reassignments. In the CS strategy, we can use such a reassignment scheme at

4. N UMERICAL R ESULTS AND D ISCUSSION

Blo c king Pro bability (%)

14 12 10

CP H3R H3 CS R

8

CS

6 4 2 0 30

40

50

60

70

80

90

traffic lo ad

Fig. 3. Total blocking per scheme for rate distribution 40:30:20:10 0.9

Co de Utilizatio n

the arrival of new calls to completely eliminate code blocking. We will refer to this strategy as the complete sharing with reassignments (CSR). The situation is slightly different in the HP strategies, where mixing of higher and lower layer codes is possible only within the common code areas. For example in the HP1 strategy, at the arrival of a iR call that finds the partition iR full, reassignments may occur in the following cases: 1) if a 2iR call is serviced by some 2iR code in the partition of iR calls, and this 2iR call can be reassigned to the partition of 2iR calls. 2) if the iR call can only be accommodated in the partition of iR/2 calls by reassigning one or two iR/2 calls. This case may cause further reassignments. We will refer to this strategy as the HP1 with reassignments (HP1R).

0.85

CP

0.8

H3R

0.75 0.7 0.65

H3 CS R CS

0.6 0.55 0.5 0.45

The performance of the proposed schemes is evaluated through event driven simulation on a 9-layer OVSF code tree. The capacity of the tree is 256R. Calls are assumed to arrive according to a Poisson process, while their call duration is exponentially distributed with equal mean. The possible rates for a new call are R, 2R, 4R, and 8R, each with a different probability of appearance. The uniform rate distribution (R:2R:4R:8R = 25:25:25:25), and a traffic pattern with lower rates being more probable (R:2R:4R:8R=40:30:20:10) are used in our simulations. Under the uniform rate distribution, the search for the most fair partition of codes yielded (n1 , n2 , n4 , n8 ) = (18, 17, 17, 17), while under the non-uniform rate distribution, (n1 , n2 , n4 , n8 ) = (36, 28, 19, 11) was chosen as the optimal partitioning at total incoming traffic loads ranging from 80 to 120. To ensure the stability of the results, each simulation run consisted of at least 100000 incoming calls. We studied two performance measures, the blocking probability P per traffic class and the utilization of codes, which is defined as

all accepted calls

(call rate)×(call duration)

. As shown in Figure 3, CS and CSR perform much better than CP, HP3 and HP3R, especially as the load increases. However, this benefit is not proportionally reflected at the utilization of the code tree as we can see in Figure 4. At light load, all the schemes utilize the codes equally, but as the load increases CSR achieves slightly better utilization than the others. The schemes listed in decreasing order of total code utilization: CSR, CS, HP3R, HP3 and CP. Although CSR achieves higher total utilization, this is not the case for each traffic class when examined separately. Lower rate codes have far better utilization than higher rate codes and that becomes more intense as the load increases. (256R)×(total simulation time)

0.4 0.35 40

50

60

70

80

90

traffic lo ad

Fig. 4. Total code utilization per scheme for rate distribution 40:30:20:10

In other words, CS and CSR accept the greatest number of calls (low total blocking), but it is clear that they favor low rates as the load increases. Figures 5 and 6 show that CSR is better in terms of fairness than the CS policy, which is due to the elimination of code blocking that mostly affects higher rates. However, at increasing loads even CSR cannot guarantee fair access. In fact at medium to heavy load, the high rates calls are practically excluded from service. Figures 7, 8 and 9, show the effects of our proposed code reservation schemes. CP treats high and low rate calls similarly at any traffic load condition. The main drawback of CP is the higher blocking under light loads. With HP3 we achieve lower blocking than CP at light loads, but certainly higher than CS and CSR, while we manage to preserve more fair access at medium loads. At medium loads, HP3R manages to keep or even strengthen the fairness property of CP while it keeps the blocking per rate even lower than HP3. Finally, we kept the same code partitioning and tested CP, HP and HPR under small variations of the incoming traffic rate distributions. Due to space limitations, we only show the performance of the HP3 scheme (Figure 10) when the partitioning of the code tree is not altered ((n1 , n2 , n4 , n8 ) = (36, 28, 19, 11)) and the rate distribution is changed to (R:2R:4R:8R=38:28:22:12). The results confirmed our intuition that HP and HPR schemes, are capable of absorbing such small traffic fluctuations, while CP is

18

8R 4R

Blo cking Pro bability (%)

Blo cking Pro bability (%)

24 22 20 18 16 14 12 10 8 6 4 2 0

2R 1R

16

8R

14

4R

12

2R

10

1R

8 6 4 2

30

40

50

60

70

80

0

90

30

traffic load

40

50

60

70

80

90

traffic load

Fig. 5. Blocking per rate for CS and rate distribution 40:30:20:10

Fig. 7. Blocking per rate for HP3 and rate distribution 40:30:20:10

12 14

4R

Blo cking Pro bability (%)

Blo cking Pro bability (%)

8R

10

2R

8

1R

6 4 2

12

8R 4R

10 8

2R 1R

6 4 2

0 30

40

50

60

70

80

90

0

traffic load

30

40

50

60

70

80

90

traffic load

Fig. 6. Blocking per rate for CSR and rate distribution 40:30:20:10 Fig. 8. Blocking per rate for HP3R and rate distribution 40:30:20:10

[1] A.C. Kam, T.Minn, and K.-Y.Siu. Dynamic assignment of orthogonal variable-spreading-factor codes in W-CDMA. IEEE Journal on Selected Areas in Communications, 18(8):1429–1440, Aug 2000. [2] Yu-Chee Tseng, Chih-Min Chao, and Shih-Lin Wu. Code placement and replacement strategies for wideband CDMA OVSF code tree management. In Proc. of IEEE GLOBECOM, volume 1, pages 562–566, 2001. [3] A. Rouskas and D. Skoutas. OVSF code assignment and reassignment at the forward link of W-CDMA 3G systems. In Proc. of IEEE PIMRC 2002. [4] P. Bahl, I. Chlamtac, and A. Farago. Resource assignment for integrated services in wireless ATM networks. International Journal of Communication Systems, 11:29–41, 1998. [5] W.-T.Chen, Y.-P.Wu, and H.-C.Hsiao. A novel code assignment scheme for W-CDMA systems. In Proc. of 54th IEEE Vehicular Technology Society Conference, volume 2, pages 1182–1186, 2001. [6] A.C.Kam, T.Minn, and K.-Y.Siu. Supporting rate guarantee and fair access for bursty data traffic in W-CDMA. IEEE Journal on Selected Areas in Communications, 19(11):2121–2130, Nov 2001. [7] F.Adashi, M.Sawahashi, and K.Okawa. Tree structured generation of orthogonal spreading codes with different lengths for forward link of DS-CDMA mobile radio. Electronic Letters, 33:27–28, Jan 1997. [8] 3GPP. Technical Specification TS 25.213 V4.2.0, Spreading and Modulation (FDD). Dec. 2001.

14 Blo cking Pro bability (%)

R EFERENCES

16

12

8R 4R 2R

10 1R

8 6 4 2 0 30

40

50

60

70

80

90

traffic load

Fig. 9. Blocking per rate for CP and rate distribution 40:30:20:10

22 Blo cking Probability (%)

less immune and loses its fairness property. The suprising result was that HP3 performed better than HP3R. This behavior is mainly explained by the inherent property of the reassignments which attempt to preserve the initial partitioning, although this is not proper for the new rate distribution.

20

8R

18 16

4R

14

2R

12

1R

10 8 6 4 2 0 30

40

50

60

70

80

90

traffic load

Fig. 10. Blocking per rate for HP3 and rate distribution 38:28:22:12