Efficient resource allocation for China's 3G/4G wireless ...


Published on

Published in: Technology, Business
  • Be the first to comment

  • Be the first to like this

No Downloads
Total Views
On Slideshare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Efficient resource allocation for China's 3G/4G wireless ...

  1. 1. WIRELESS COMMUNICATIONS IN CHINA: TECHNOLOGY VS. MARKETS Efficient Resource Allocation for China’s 3G/4G Wireless Networks Yu Cheng, University of Toronto Hai Jiang and Weihua Zhuang, University of Waterloo Zhisheng Niu and Chuang Lin, Tsinghua University, China ABSTRACT techniques with efficient resource utilization to support real-time multimedia (video/audio) ser- The all-IP DiffServ model is expected to be vices and non-real-time data services, respective- the most promising architecture for QoS provi- ly, in 3G/4G wireless networks. sioning in China’s next-generation wireless net- It is very likely that the three 3G air interface works, due to its scalability, convenience for models, wideband code-division multiple access mobility support, and capability of interworking (WCDMA), CDMA2000, and the homegrown heterogeneous radio access networks. This arti- time-division synchronous CDMA (TD- cle focuses on efficient resource allocation in a SCDMA), will all be deployed in China. An all- wireless DiffServ architecture. Resource utiliza- IP differentiated services (DiffServ) platform tion efficiency is particularly important for [1–3] is the most promising architecture to inter- China’s wireless networks as the mobile user work the heterogeneous wireless access networks density in China is and will continue to be much and the Internet to provision broadband access, higher than that in other countries. More specifi- seamless global roaming, Internet/telecommuni- cally, we propose a novel buffer sharing scheme cation services anywhere, and QoS guarantee for to provide assured service for real-time layer- various IP multimedia services. The reasons are coded multimedia traffic, which can guarantee as follows. First, DiffServ [4] is a scalable class- the specific packet loss requirement of each based traffic management mechanism without layer with UDP as the transport layer protocol. using per-flow resource reservation and per-flow An adaptive optimal buffer configuration can be signaling in the core routers; second, DiffServ applied to achieve maximum resource utilization adopts a domain-based resource management over the time-varying channel. Assured service is model. Each domain can freely choose whatever also provided to TCP data traffic for guaranteed mechanism is proper for internal resource man- throughput, where the cross-layer coupling agement as long as its service level agreements between the TCP layer and link layer is exploit- (SLAs) with neighboring domains are met. Such ed to efficiently utilize the wireless resources. a domain-based architecture is very convenient for the interconnection of heterogeneous wire- INTRODUCTION less networks [1]. Lastly, a domain-based archi- tecture can be seamlessly integrated with After many indoor and outdoor tests, extensive micromobility protocols to support fast handoff upgrade of network infrastructures, and pre- [3]. This advantage can considerably benefit market promotion of new services, China will China’s 3G/4G wireless networks where micro/ deploy commercial third-generation (3G) picocellular deployment is required for high mobile/wireless networks very soon. To build up resource utilization. As a result, this article a stable and profitable market for 3G wireless focuses on efficient resource allocation in an all- communications, China has to pay particular IP DiffServ wireless architecture. attention to efficient management of wireless The 3G/4G wireless systems are designed for resources, because its 3G wireless networks have multimedia communications, where broadband to support mobile users with much higher densi- mobile video services will become reality for the ty than that in other countries. The requirement first time. Recent advances in video coding have for efficient resource allocation will persist in made it possible to encode video with a very future-generation wireless networks, due to the flexible layering structure, where a base layer limited radio spectrum, huge number of mobile contains most important features of the video subscribers, and high transmission rate require- and some enhancement layers contain data for ments of new multimedia services. In this article refining the reconstructed video quality. The we propose two novel quality of service (QoS) layer coding concept can also be applied to 76 0163-6804/05/$20.00 © 2005 IEEE IEEE Communications Magazine • January 2005
  2. 2. audio traffic. Such a layered structure supports adaptive multimedia services with different bandwidth requirements by adjusting the num- DiffServ ber of layers delivered. In wireless networks, this TD-SCDMA type of adaptive services is very important for domain DiffServ efficient resource utilization, because wireless WCDMA DiffServ domain resource availability fluctuates due to user mobil- WCDMA ity and time-varying channel quality. In the fol- domain lowing, we propose a novel buffer sharing scheme to provide an assured service to the layer-coded multimedia traffic, which can guar- DiffServ antee the specific packet loss requirement of Internet each layer with User Datagram Protocol (UDP) as the transport layer protocol for real-time ser- DiffServ Core vices. Adaptive optimal buffer configuration is router wireless LAN applied to achieve maximum resource utilization DiffServ CDMA2000 over the time-varying channel. For data traffic domain Edge router with Transmission Control Protocol (TCP) at the transport layer, the assured service is provi- Gateway router sioned for a guaranteed throughput. As in wire- SLA negotiated less networks TCP performance is heavily affected by the loss due to unstable wireless channel quality, we propose a cross-layer model to capture the coupling between the TCP trans- Air interface Base port layer and the link layer. The target TCP statoin throughput is guaranteed with minimal resources required at the link layer and therefore at the physical layer. DiffServ Mobile wireless domain station ALL-IP DIFFSERV ARCHITECTURE (3G/4G access) The all-IP DiffServ architecture under consider- ation is shown in Fig. 1, which enables access to Internet/telecommunication services indepen- dent of the air interface technique. In the fol- lowing, we use Internet services to generally represent all the possible Internet/telecommuni- n Figure 1. An all-IP DiffServ architecture for 3G/4G wireless communica- cation services that may appear in the all-IP tions. wireless/wireline networks. In the all-IP DiffServ architecture, a number of nearby radio access networks (RANs) having the same air interface routers use several separate queues, controlled are grouped into a wireless DiffServ domain, by certain scheduling algorithms, to provide dif- and all the domains are connected through the ferentiated classes of services. DiffServ Internet backbone to provide end-to- end Internet services to a mobile station (MS). SERVICE CLASS MAPPING IN Although wireless local area networks (WLANs) WIRELESS DIFFSERV can be seamlessly integrated to the all-IP archi- tecture, here we focus on resource allocation in Among the 3G standards, the Universal Mobile cellular systems. Telecommunications System (UMTS) based on In each DiffServ wireless domain, the RAN WCDMA technologies defines four QoS classes: architecture is the same as that defined in the conversational, streaming, interactive, and back- 3G specifications, but all the network elements ground [2]. Conversational and streaming classes are enhanced to fulfill the functions of a Diff- are intended for real-time traffic. They both pre- Serv IP router. The gateway and base stations serve time relations between adjacent informa- are edge routers of the domain and connected tion elements of the stream, but conversational through core routers. The gateway is the inter- class has stricter delay requirements. For the face to the DiffServ Internet backbone. For interactive and background classes, transfer example, the gateway General Packet Radio Ser- delay is not a main concern, but error-free deliv- vice (GPRS) support node (GGSN) in the ery of data should be guaranteed. The UMTS WCDMA domain is the gateway of the domain QoS class definitions can also be extended to to the external DiffServ Internet. In the gateway, other 3G/4G wireless domains. On the other SLAs are negotiated to specify the resources hand, DiffServ defines the expedited forwarding allocated by the Internet service provider to (EF) [5] per-hop behavior (PHB) for premium serve the aggregate traffic flowing from/into the service and the assured forwarding (AF) PHB domain. The gateway conditions the aggregate [6] for assured service, in addition to the classic traffic for each service class according to the best effort service. To extend IP services to the SLA resource commitments. The base stations wireless domain, the UMTS QoS classes must be provide MSs with access points to the Internet, mapped to the DiffServ classes [2]. and perform per-flow traffic conditioning and Normally, the conversational class can be marking for uplink transmission. All DiffServ mapped to the EF PHB for very low-delay and IEEE Communications Magazine • January 2005 77
  3. 3. low-loss service, streaming and interactive traffic ciently reduce interference and improve the suc- The all-IP to the AF PHB, and background traffic to best- cessful transmission rate of data packets. In such effort service [2]. However, the peak rate band- a scheme, the SINR value for each MS should architecture width allocation for EF services is not suitable be designed carefully. provides an efficient for wireless communications. The constant (peak For real-time multimedia traffic, UDP is usu- way to serve rate) bandwidth requirement during each session ally used in the transport layer to avoid retrans- lifetime is very hard to guarantee due to the mission delay, while the link and physical layers multimedia traffic time-varying wireless channel capacity. Further- try to provide stable capacity to upper layers by over heterogeneous more, Internet traffic often has a bursty charac- assigning a large SINR and/or powerful FEC for teristic; inefficient peak rate allocation is reliable transmission. For non-real-time data wireless domains. particularly undesirable for wireless transmission traffic, TCP is usually adopted for an error free For each wireless with scarce resources. In DiffServ, AF PHBs aim connection. TCP dynamically adjusts its conges- domain, the at providing throughput guaranteed services with tion window according to the network conges- in-flow loss differentiation. When combined with tion status (e.g., the packet loss event rate and bandwidth admission control, AF PHBs can be extended to round-trip delay) and retransmits the lost TCP requirement from a very flexible service architecture. Taking into segments. Over a wireless link, the link layer account that current video/audio codecs can resource allocation ultimately determines the the network layer work properly with a small packet loss probabili- transmission delay and packet loss event rate, will eventually be ty, we propose that all the real-time classes (i.e., and therefore affects TCP performance. As dis- mapped to a conversational and streaming classes) and non- cussed below, in the DiffServ architecture a tar- real-time class (i.e., interactive class) are mapped get TCP throughput is guaranteed for wireless resource into AF PHBs to guarantee statistical QoS non-real-time data traffic, and a cross-layer opti- requirement at the requirements and improve resource utilization. mization between the transport and link layers is Specifically, we propose using a partitioned first- proposed for efficient resource utilization over link layer. in-first-out (FIFO) buffer [7] to provide an AF the wireless link. PHB for real-time layer-coded multimedia traffic under UDP. For non-real-time TCP data traffic, an AF PHB is provisioned to guarantee a speci- BUFFER SHARING FOR LAYER-CODED fied throughput, and a cross-layer model is pro- posed to capture the coupling between the MULTIMEDIA TRAFFIC transport and link layers to achieve efficient Providing real-time multimedia services to wireless resource utilization. QoS provisioning mobile users is one of the prominent features of for real-time and non-real-time AF PHBs will be 3G/4G wireless communications. Layer-coded discussed in separate sections, after some gener- video/audio traffic is very suitable for transmis- al principles for wireless resource allocation over sion over the wireless link. We propose a novel IP-based 3G networks are described. buffer-sharing approach to serve the layered multimedia traffic. Consider video services as an WIRELESS RESOURCE ALLOCATION example. Suppose a video source generates traf- Based on CDMA technology, all three 3G stan- fic having J (J ≥ 1) layers, referred to as J classes. dards are likely to be deployed in China. The all- The QoS requirement is specified by a packet IP architecture provides an efficient way to serve loss probability (PLP) εj for class j, j ∈ {1, 2, …, multimedia traffic over heterogeneous wireless J}. Letting class J represent the base layer infor- domains. For each wireless domain, the band- mation and have the most stringent QoS require- width requirement from the network layer will ment, we have ε 1 > ε 2 > … > ε J > 0. A buffer eventually be mapped to a wireless resource of size B serves traffic with a channel capacity c requirement at the link layer. provided by the link layer, where parameters B For circuit-switched voice traffic the resource and c determine the queuing delay bound at the allocation is relatively simple, and the 2G buffer. The traffic admission policy is based mobile/wireless systems have shown success in upon a buffer space reservation scheme, using a serving voice applications over wireless channels. buffer partition vector Bt = (B1, B2, …, BJ–1) to However, for IP-based multimedia traffic, the provide J loss priorities, where 0 = B 0 < B 1 < case is quite different where different applica- B2 < … < BJ–1 < BJ = B. Let X be the number tions have different QoS (e.g., transmission rate of packets of all the classes in the buffer at time and accuracy) requirements, and packet switch- t. When Bj–1 ≤ X < BJ (1 ≤ j ≤ J), only traffic of ing should be used to efficiently utilize the scarce classes {j, j + 1, … , J} is admitted into the radio resources. Generally, orthogonal variable buffer. Therefore, higher classes of traffic are spreading factor (OVSF) codes or multicode served with higher priority by access to a larger transmission can be employed to support high- buffer space. The partitioned buffer scheme is an rate applications; and high transmission accuracy implementation of the AF PHB defined in the can be achieved by assigning a high signal-to- DiffServ architecture. noise (plus interference)-ratio (SINR) and/or a The loss calculation and admission control in powerful forward error correction (FEC) code, the partitioned buffer system are required to or by means of automatic repeat request (ARQ). serve multiplexed layer-coded video sources with It is shown in [8] that, in the uplink, the smaller QoS guarantee; that is, the actual PLP experi- bit error rate an MS requires, the higher SINR enced by class j traffic (PLPj) should satisfy PLPj the MS should receive, and the more resources ≤ εj (1 ≤ j ≤ J). The loss analysis is relatively sim- should be allocated to the MS to transfer a link ple when the buffer has a large size (the delay layer packet. As CDMA systems have interfer- requirement is not strict). For a large buffer, it is ence limited capacity, a properly designed power widely agreed that the packet loss probability allocation and access control scheme can effi- can be well approximated by the overflow proba- 78 IEEE Communications Magazine • January 2005
  4. 4. bility in an infinite buffer system, that is, PLPj ≈ loss probability in a finite buffer system with size P {X > Bj} (1 ≤ j ≤ J). Furthermore, when each B can be accurately estimated from the overflow When the wireless partition region B j – B j–1 is large enough, the probability P{X > B} of an infinite buffer sys- queuing process in neighboring regions can be tem, according to a simple mapping relation channel capacity approximately considered independent of each PLP ≈ αP{X > B}, where α < 1 is a constant. varies with time, a other, if the input traffic is short-range depen- We extend the mapping relation to the parti- layer-coded source dent (SRD).1 Based on the above facts, PLPs of tioned buffer (with finite size) case by proving the J traffic classes in the partitioned buffer can that PLPj ≈ αjP{X > Bj} for all J classes of traf- can adjust the be approximated by a two-step algorithm: fic. Another foundation of the proposed tech- bandwidth 1)Calculate PLP 1 ≈ P{X > B 1 }, using the nique is the calculation technique of overflow large buffer overflow approximation. probability for FBM traffic, based on which we requirement 2)Calculate PLP for classes 2 to J iteratively, find the relation P{X > B j } ≈ f(P{X > B j–1 }, correspondingly as PLP j ≈ P{X > B j} ≈ P{X > B j–1} ⋅ P{X P{X > B j – B j–1 }, H), where the function f(⋅) by increasing or > B j – B j–1} with overflow approximation captures the correlation between the queuing applied in each region Bj – Bj–1. The itera- processes of neighboring partition regions by decreasing the tive calculation has an explicit physical taking the Hurst parameter H into account. All number of layers meaning, (i.e., the overload at B j happens the derivations and related references are given in two steps): the first is that the queue in [9]. By adopting the new techniques, the loss transmitted for a occupancy goes into the jth partition region, calculation in a finite-size partitioned buffer is as better tradeoff and the second is that an overload event follows: between efficient happens in that region. Note that the traffic 1. Model or substitute the input traffic with an arrival process in each region is different FBM process. resource utilization due to the buffer sharing policy. 2. Calculate P{X > B1} using the overflow cal- and QoS Details of the large buffer analysis can be culation technique for the FBM process. found in our study [7]. When the traffic statisti- 3. Iteratively calculate P{X > B j} ≈ f(P{X > satisfaction. cal parameters and buffer size are fixed, the B j–1 }, P{X > B j – B j–1 }, H) for 2 ≤ j ≤ J, choice of B t determines the channel capacity where the overflow calculation technique required to guarantee the QoS for all J classes. for FBM is applied in each region Bj – Bj–1. The selection of an optimal partition vector B* t 4. Find the PLP according to the mapping results in the minimal channel capacity require- relation as PLPj ≈ αjP{X > Bj}. ment c*, in other words, the maximum resource The optimal buffer partitioning concept can utilization. An algorithm to jointly solve B* and t also be applied to the finite buffer FBM case for c* is given in [7]. The capacity c* can be defined maximum resource utilization. The loss analysis as the minimal effective bandwidth and used to based on FBM can not only deal with the finite- achieve linear admission control. Linear admis- size effect, but also achieve a statistical multi- sion control is desired for online operation. It plexing gain by considering the aggregate traffic avoids directly solving the optimization problem directly (to be demonstrated by numerical exam- for the aggregate traffic, by approximating the ples). minimal resource requirement of the aggregate The proposed partitioned buffer technique traffic with the summation of the minimal effec- can be applied to all layer-coded multimedia tive bandwidths. The optimal partition vector for traffic. When the wireless channel capacity varies the admitted traffic can be heuristically approxi- with time, a layer-coded source can adjust the mated by a weighted summation of the single- bandwidth requirement correspondingly by source optimization results. increasing or decreasing the number of layers To serve the real-time conversational/stream- transmitted for a better trade-off between effi- ing multimedia traffic, the buffer size should be cient resource utilization and QoS satisfaction. small to achieve a low queuing delay. The large The buffer configuration can be adjusted corre- buffer technique performs poorly in this case. spondingly to serve the traffic. An example of Furthermore, it is quite possible that the input such adaptive service and the procedure to traffic may be long-range dependent (LRD). To dynamically configure the partitioned buffer can address these two problems, we develop a loss be found in our study given in [3]. calculation technique for a finite-size partitioned buffer with fractional Brownian motion (FBM) input. FBM is a Gaussian process with stationary CROSS-LAYER OPTIMIZATION FOR increments, characterized by three parameters: the mean arrival rate λ, the variance of traffic in TCP DATA TRAFFIC a time unit σ 2 , and the Hurst parameter H While UDP is often used as the transport layer describing the correlation in the traffic arrival protocol for real-time multimedia service, TCP process. FBM is a model originally used to char- should be used for error-free delivery of data acterize LRD (self-similar) Internet traffic. traffic. Over an air interface, the coupling However, recent studies show that the aggregate between the TCP and link layers should be con- of SRD sources can be equivalently substituted sidered in order to efficiently utilize the scarce by an FBM process, from the perspective that wireless resources as well as satisfy the through- either the buffer overflow probability or avail- put requirement. In our system, a TCP segment able capacity is preserved after substitution. consists of a number of link layer packets to be 1 The Markov modulated Therefore, the loss calculation technique based ready for transmission over the wireless link. fluid source or Markov on FBM traffic is a general technique that can TCP was originally developed for wireline modulated Poisson pro- be applied to both LRD and SRD input sources. networks with a reliable physical layer, where cess are two classic short- One foundation of the proposed loss calcula- packet loss mainly results from network conges- range dependent models tion technique is the recent observation that the tion. In a wireless environment, TCP perfor- of video traffic [7]. IEEE Communications Magazine • January 2005 79
  5. 5. layer packets, the bandwidth is not utilized effi- 100 ciently if such a packet discarding event occurs after the main part of the TCP segment has been transmitted successfully. Hence, here we propose unlimited retransmissions (until success- 10–1 ful receipt) to improve resource utilization for delay-tolerant data traffic. As retransmissions use part of the bandwidth, TCP will then see a 10–2 reliable link with larger delay and smaller band- Loss probability width. A concern about unlimited retransmissions is 10–3 possible performance degradation due to com- peting retransmissions between TCP and the link layer ARQ. TCP Reno and its variants usu- 10–4 ally employ coarse granularities of retransmis- sion timeout value (typically multiples of 500 Class 1, simulation ms). Hence, the impact of competing retrans- Class 1, finite buffer analysis 10–5 Class 1, large buffer approximation missions on TCP performance is not significant Class 2, simulation [11]. It is also shown in [12] that an unlimited Class 2, finite buffer analysis Class 2, large buffer approximation retransmission mechanism is effective in terms 10–6 of TCP throughput performance and wireless 0 10 20 30 40 50 60 70 80 energy savings. Furthermore, a random early Partition threshold B1 detection (RED) buffer rather than a drop-tail n Figure 2. Loss probabilities of two-layer coded traffic with B = 80 packets buffer is employed at the wireless sender side to control the possible delay inflation caused by and c = 340 packets/s. unlimited retransmissions. RED has the ability to control the queuing delay and prevent corre- lation of packet dropping events (i.e., consecu- mance can degrade severely due to unreliable tive packet losses). Each TCP flow has its own link layer transmission. In order to reduce the RED buffer for the wireless link. As the per- link layer packet loss rate seen by TCP, the base flow scheme applies only at the edge of the station employs a radio link protocol (RLP) DiffServ domain, it will not incur any scalability which uses a selective repeat automatic repeat problem. In addition to the RED buffer, each request (SR-ARQ) error recovery mechanism to flow (say, flow i) over the wireless link also retransmit link layer packets not received suc- keeps a link layer buffer for transmission and cessfully in the previous frames. Current RLP retransmission of link layer packets. The two protocols normally have a constraint on the buffers can be viewed as a virtual RED buffer at number of retransmissions (i.e., if a link layer the link layer. To serve this virtual buffer, up to packet cannot be received successfully after a M i link layer packets from TCP flow i can be given number of retransmissions, the sender will scheduled (by multicode CDMA transmissions) discard this packet and the subsequent link layer in each radio frame (i.e., link layer time frame) packets belonging to the same TCP segment not with SINR value Γi, where (Mi, Γi) is referred to transmitted yet). This is important to control the as the resource vector. transmission latency in a correlative fading chan- It has been shown that TCP throughput T is nel where bursty packet losses may be induced determined by the packet loss event rate and due to inaccurate power control. We have shown round-trip delay [13]. Assuming that a TCP flow that in the link layer a delayed multiple copy i experiences a fixed delay and a fixed packet retransmission (DMCR) scheme under a maxi- loss event rate in the wireline part, its through- mum number of retransmissions can benefit put is determined mainly by the performance of higher-layer TCP performance [10]. However, as the wireless link, which is further determined by a TCP segment normally consists of many link the wireless resource vector (Mi, Γi). To achieve (K5, K6) (20, 0) (16, 2) (12, 4) (8, 6) (4, 8) (0, 10) Optimization with * B1 821 801 779 754 727 696 large buffer c* 1369.2 1369.2 1369.2 1369.2 1369.2 1369.2 approximation PLP1 2.04 × 10–3 1.35 × 10–3 1.07 × 10–3 9.97 × 10–4 9.20 × 10–4 8.09 × 10–4 and simulated PLP PLP2 1.10 × 10–6 8.90 × 10–7 8.33 × 10–7 5.47 × 10–7 4.67 × 10–7 2.19 × 10–6 Optimization with * B1 946 912 878 848 823 805 finite buffer c* 1312.9 1279.1 1248.4 1220.0 1193.3 1167.8 analysis and PLP1 5.43 × 10–3 5.68 × 10–3 5.29 × 10–3 4.71 × 10–3 4.88 × 10–3 5.05 × 10–3 simulated PLP PLP2 6.25 × 10–5 3.49 × 10–5 3.87 × 10–5 4.14 × 10–5 7.10 × 10–5 6.71 × 10–5 n Table 1. Optimal buffer configurations and simulated packet loss probabilities: large buffer approxima- tion vs. finite buffer analysis. 80 IEEE Communications Magazine • January 2005
  6. 6. the target TCP throughput, for each possible Mi value the Γi value is determined iteratively. For 200 an intermediate Γi value, we calculate the wire- Analytical less link delay and packet loss event rate, and Simulation 190 further estimate the achieved TCP throughput T. If T is greater (less) than the required through- put, we decrease (increase) the Γ i value. This 180 procedure is repeated until the achieved T con- verges to the target throughput, and we call the 170 obtained (Mi, Γi) a feasible resource vector. There Achieved TCP throughput (kb/s) may exist many feasible resource vectors for a 160 target throughput. We define the normalized resource amount required by flow i in a radio frame as miΓi/(G + miΓi) [8], where mi (≤ Mi) is 150 the number of actually scheduled link layer packets in the radio frame and G the CDMA 140 processing gain. The average normalized resource amount over all the radio frames is termed the equivalent resource amount. Among 130 all the feasible resource vectors, we can get the optimal one that minimizes the equivalent 120 resource requirement at the link layer, therefore achieving the maximal resource utilization. The 110 mapping between the required throughput and the optimal resource vector is obtained at the 100 connection setup via a table lookup. More 20 22 24 26 28 30 32 detailed discussion of the cross-layer optimiza- Number of scheduled packets per frame Mi tion is given in [8]. n Figure 3. Achieved TCP throughput vs. Mi for a data flow with a throughput PERFORMANCE EVALUATION requirement of 150 kb/s. In this section analysis and computer simulation results are presented to demonstrate the perfor- mance of the proposed resource allocation tech- the simulation and finite buffer analysis results niques. We first show the resource utilization are in a very close match, while the large buffer improvement by optimally configuring a parti- approximation deviates far from the simulation tioned buffer for real-time traffic, and then results, especially for the class 2 traffic where the demonstrate the performance of the cross-layer finite buffer effect is more obvious. Accurate optimization for non-real-time data traffic. loss analysis in the small buffer case is particu- larly useful for real-time traffic having both OPTIMAL BUFFER CONFIGURATION WITH delay and loss requirements. ACCURATE LOSS CALCULATION The accurate loss calculation technique can be used in admission control to further improve For a large buffer, the effectiveness of the resource utilization. Here we re-examine the optimal buffer partitioning and minimal effec- admission control example considered in [7] by tive bandwidth to improve resource utilization substituting the input traffic with the FBM pro- has been demonstrated in [7]. Here, we show cess and applying the accurate loss calculation that resource utilization can be further technique instead of the overflow approxima- improved by accurately calculating the PLP tion. Heterogeneous two-class on/off sources instead of using overflow approximation. The (type-5 and type-6 sources in [7]) are consid- units of traffic arrival rate and channel capacity ered, with the PLP requirement of 10 –2 and are packet per second, and the unit of buffer 10–4 for class 1 and class 2, respectively. Specifi- size is packet. cally, type-5 sources with the minimal effective The first example demonstrates the accuracy * bandwidth of c 5 = 68.46 and type-6 sources of the proposed loss calculation technique for a with c* = 2c* are multiplexed in a partitioned 6 5 partitioned buffer of finite size. Consider a two- buffer of size B = 1000 and served with chan- class FBM input served with a partitioned buffer nel capacity c = 20c * . With K 5 (K 6 ) denoting 5 of size 80. The channel capacity c is 340. In this the number of accepted type-5 (type-6) sources, setting, the worst case delay is 80/340 ≈ 235 ms, Table 1 lists the optimal partition threshold B* 1 small enough for the streaming class services. and the minimal channel capacity to guarantee The FBM source consists of a class-1 (enhance- QoS, obtained from large buffer approximation 2 ment layer) FBM with λ 1 = 20, σ1 = 50, and a and FBM finite buffer analysis, respectively. 2 class-2 (base layer) FBM with λ 2 = 300, σ 2 = The simulation results for loss probability cor- 500, with the Hurst parameter H = 0.83 for both responding to the two configurations are also classes. The buffer is partitioned into two regions given in Table 1. It can be observed that the to differentiate the loss behaviors of class 1 and minimal effective bandwidth and linear admis- class 2. Figure 2 shows the loss probabilities of sion control from large buffer approximation two classes vs. the partition threshold B 1 , are a conservative resource allocation approach, obtained from analysis and computer simulation. over-satisfying the QoS; on the other hand, the It is observed that in this small buffer example, FBM finite buffer analysis technique clearly IEEE Communications Magazine • January 2005 81
  7. 7. TCP flow. An optimal point (M i = 21 in this 1 example) can be found among the feasible Analytical Simulation resource vectors for the minimal equivalent 0.9 resource requirement, thus achieving maximal resource utilization. 0.8 CONCLUSION Required equivalent resource amount 0.7 The prominent characteristics of the 3G/4G wireless communications in China are coexis- 0.6 tence of heterogeneous access technologies and high distribution density of mobile users. This article addresses efficient resource man- 0.5 agement of such a 3G/4G wireless network by suggesting an all-IP DiffServ wireless architec- 0.4 ture. In particular, we present two novel QoS techniques for efficient resource allocation. 0.3 One contribution is a partitioned buffer scheme that provisions UDP layer-coded multimedia traffic via an AF PHB with delay and packet 0.2 loss guarantee. Maximum resource utilization is achieved by accurate loss analysis in the par- 0.1 titioned buffer, exploitation of statistical multi- plexing through FBM modeling, and adaptive 0 optimal buffer configuration for the time-vary- 20 22 24 26 28 30 32 ing wireless channel. The other contribution is Number of scheduled packets per frame Mi the implementation of an AF PHB to serve TCP data traffic with a guaranteed throughput. n Figure 4. Equivalent resource requirement vs. Mi for a data flow with a To improve TCP performance over the air interface, the coupling between the TCP and throughput requirement of 150 kb/s. link layers is considered, and a cross-layer opti- mization technique is proposed for efficient utilization of wireless resources. Although this improves resource utilization by more accurate study is motivated by the stringent requirement calculation of the loss probability in a finite for efficient resource allocation in China’s 3G/ buffer and better exploitation of statistical mul- 4G wireless communications, the proposed tiplexing among multiple flows. QoS and resource allocation techniques can be applied to general 3G/4G wireless networks. PERFORMANCE OF CROSS-LAYER OPTIMIZATION Consider a single data flow initiated from an MS with a target throughput requirement of 150 kb/s REFERENCES in a single-cell CDMA system. The radio frame [1] B. Moon and H. Aghvami, “DiffServ Extension for QoS length is set to be 20 ms. Each link layer packet Provisioning in IP Mobility Environments,” IEEE Wireless Commun., vol. 10, no. 5, Oct. 2003, pp. 38–44. has 192 bits, of which 160 bits are payload. The [2] S. I. Maniatis, E. G. Nikolouzou, and I. S. Venieris, “QoS processing gain G for each link layer packet is Issues in the Converged 3G Wireless and Wired Net- 128. In the transport layer, a TCP segment has a works,” IEEE Commun. Mag., vol. 40, no. 8, Aug. 2002, fixed size of 8000 bits. To achieve the target pp. 44–53. [3] Y. Cheng and W. Zhuang, “DiffServ Resource Allocation throughput requirement, we first obtain the fea- for Fast Handoff in Wireless Mobile Internet,” IEEE sible resource vectors (M i , Γ i ), calculate the Commun. Mag., vol. 40, no. 5, May 2002, pp. 130–36. equivalent resource amount required by each [4] S. Blake et al., “An Architecture for Differentiated Ser- vector, and then run computer simulations to vices,” IETF RFC 2475, Dec. 1998. [5] V. Jacobson, K. Nichols, and K. Poduri, “An Expedited examine the selected feasible resource vectors. Forwarding PHB,” IETF RFC 2598, June 1999. For each selected feasible resource vector, 10 [6] J. Heinanen et al., “Assured Forwarding PHB Group,” independent simulations are carried out. Each IETF RFC 2597, June 1999. simulation runs for 30, 000 radio frames, and the [7] Y. Cheng and W. Zhuang, “Effective Bandwidth of Mul- ticlass Markovian Traffic Sources and Admission Control statistics are collected in the last 24, 000 frames. with Dynamic Buffer Partitioning,” IEEE Trans. Com- In the simulations, the achieved throughput is mun., vol. 51, no. 9, Sept. 2003, pp. 1524–35. traced through the accumulative Acknowledg- [8] H. Jiang and W. Zhuang, “Quality-of-Service Provision- ments (ACKs) for TCP segments from the cor- ing to Assured Service in the Wireless Internet,” Proc. IEEE GLOBECOM, vol. 6, Dec. 2003, pp. 3078–82. respondence node. [9] Y. Cheng and W. Zhuang, “Calculation of Loss Probabil- Figures 3 and 4 show the achieved TCP ity in a Partitioned Buffer with Self-Similar Input Traf- throughput and required equivalent resource fic,” Proc. IEEE GLOBECOM, 2004. amount for different feasible resource vectors, [10] Z. Niu, Y. Wu, and J. Zhu, “A Delayed Multiple Copy Retransmission Scheme for Data Communication in respectively. As the feasible resource vector has Wireless Networks,” J. Commun. and Net., vol. 5, no. 3, only one degree of freedom under the constraint Sept. 2003, pp. 222–29. of the target throughput, only values of M i are [11] H. Balakrishnan et al., “A Comparison of Mechanisms for shown in the horizontal axis. It can be seen that Improving TCP Performance over Wireless Links,” IEEE/ACM Trans. Net., vol. 5, no. 6, Dec. 1997, pp. 756–69. the simulation results closely match the numeri- [12] F. Vacirca, A. De Vendictis, and A. Baiocchi, “Investi- cal analysis results. Each feasible resource vector gating Interactions between ARQ Mechanisms and TCP can guarantee the throughput requirement of a over Wireless Links,” Proc. Euro. Wireless, Feb. 2004. 82 IEEE Communications Magazine • January 2005
  8. 8. [13] J. Padhye et al., “Modeling TCP Reno Performance: A Canada. She received the Premier’s Research Excellence Simple Model and Its Empirical Validation,” IEEE/ACM Award (PREA) in 2001 from the Ontario Government for Trans. Net., vol. 8, no. 2, Apr. 2000, pp. 133–45. demonstrated excellence of scientific and academic contri- The average butions. She is an Associate Editor of IEEE Transactions on Vehicular Technology and EURASIP Journal on Wireless normalized resource BIOGRAPHIES Communications and Networking. amount over all the YU CHENG [S’01, M’04] (y.cheng@utoronto.ca) received B.E. and M.E. degrees from Tsinghua University, Beijing, China, Z HISHENG N IU [SM] (niuzhs@tsinghua.edu.cn) graduated radio frames is in 1995 and 1998, respectively, and a Ph.D. degree from from Northern Jiaotong University, Beijing, China, in the University of Waterloo, Ontario, Canada, in 2003, all in 1985, and got his M.E. and D.E. degrees from Toyohashi termed the equiva- electrical engineering. From September 2003 to August University of Technology, Japan, in 1989 and 1992, lent resource 2004 he was a postdoctoral fellow in the Department of respectively. In 1994 he joined Tsinghua University, Bei- Electrical and Computer Engineering at the University of jing, China, where he is now a full professor in the amount. Among all Waterloo. Since September 2004 he has been a postdoc- Department of Electronic Engineering. He is also an toral fellow in the Department of Electrical and Computer adjunct professor of Beijing Jiaotong University (formally the feasible resource Engineering at the University of Toronto. His research inter- Northern Jiaotong University). His current research inter- ests include QoS provisioning in IP networks, resource ests include teletraffic theory, performance evaluation of vectors, we can get management, traffic engineering, and wireless/wireline high-speed broadband integrated networks, wireless the optimal one interworking. He received a postdoctoral fellowship from ATM/IP networks, mobile Internet, and stratospheric com- the Natural Sciences and Engineering Research Council of munication systems. He is the Chair of the Membership that minimizes the Canada (NSERC) in 2004. Development Committee of the Asia-Pacific Board of IEEE Communications Society. He is also serving as the Chair equivalent resource HAI JIANG [S’04] (hjiang@bbcr.uwaterloo.ca) received a B.S. of the Steering Committee of the Asia-Pacific Conference degree in 1995 and an M.S. degree in 1998, both in elec- on Communication (APCC) and Technical Program Chair requirement at the trical engineering, from Peking University, Beijing, China. of APCC 2004. He is a senior member of the Chinese link layer, therefore He is currently working toward a Ph.D. degree at the Uni- Institute of Electronics (CIE) and a member of the IEICE, versity of Waterloo. His current research interests include and the Beijing Representative of IEICE. achieving the QoS provisioning and resource management for multime- dia communications in all-IP wireless networks. CHUANG LIN [SM] (chlin@tsinghua.edu.cn) is a professor and maximal resource the head of the Department of Computer Science and WEIHUA ZHUANG [M’93, SM’01] (wzhuang@bbcr.uwaterloo. Technology, Tsinghua University, Beijing, China. He received utilization. ca) received B.Sc. and M.Sc. degrees from Dalian Maritime a Ph.D. degree in computer science from Tsinghua Univer- University, Liaoning, China, and a Ph.D. degree from the sity in 1994. His current research interests include comput- University of New Brunswick, Fredericton, Canada, all in er networks, performance evaluation, network security, electrical engineering. Since October 1993 she has been logic reasoning, and Petri net theory and its applications. with the Department of Electrical and Computer Engineer- He has published more than 180 papers in research jour- ing, University of Waterloo, where she is a full professor. nals and IEEE conference proceedings in these areas and She is a co-author of the textbook Wireless Communica- has published three books. He serves as Technical Program tions and Networking (Prentice Hall, 2003). Her current Co-Chair for the 10th IEEE Workshop on Future Trends of research interests include multimedia wireless communica- Distributed Computing Systems (FTDCS 2004); General tions, wireless networks, and radio positioning. She is a Chair, ACM SIGCOMM Asia Workshop 2005; and Associate licensed Professional Engineer in the Province of Ontario, Editor, IEEE Transactions on Vehicular Technology. IEEE Communications Magazine • January 2005 83