Base paper provisioning of qos

Page 1

JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 15, NO. 1, FEBRUARY 2013

61

Provisioning of QoS Adaptability in Wired-Wireless Integrated Networks Mian Guo, Shengming Jiang, Quansheng Guan, and Huachao Mao Abstract: The increasing number of mobile users and the popularity of real-time applications make wired-wireless integrated network extremely attractive. In this case, quality of service (QoS) adaptability is particularly important since some important features of the integrated network call for QoS adaptability, such as mobility, bursty applications and so on. Traditional QoS schemes include integrated service (IntServ) and differentiated service (DiffSev) as well as their variants. However, they are not able to balance well between scalability and QoS granularity. For example, IntServ faces the scalability problem, while DiffServ can only provide coarse granular QoS. In addition, they are also unable to efficiently support QoS adaptability. Therefore, a per-packet differentiated queueing service (DQS) was proposed. DQS was originally proposed to balance between scalability and QoS granularity in wired networks and then extended to wireless networks. This paper mainly discusses how to use DQS to support QoS adaptability in wired-wireless integrated networks. To this end, we propose a scheme to determine dynamic delay bounds, which is the key step to implement DQS to support QoS adaptability. Simulation studies along with some discussions are further conducted to investigate the QoS adaptability of the proposed scheme, especially in terms of its support of QoS adaptability to mobility and to bursty real-time applications. Index Terms: Differentiated queueing service (DQS), quality of service (QoS) adaptability, wired-wireless integrated networks.

I. INTRODUCTION The number of wireless users increases exponentially, and meanwhile, real-time applications become more and more popular. For example, Cisco has estimated that in the near future almost 66 percent of the mobile applications will be video [1]. Therefore, the motivation to provide integral seamless Internet services to any user anytime, anywhere and anyhow drives the integration of wired and wireless networks. Among all mechanisms required to support seamless services in this integrated network, end-to-end quality of service (QoS) support is particularly important [2], [3]. This is because the particular properties of the various access networks make it difficult to provide uniform and stable QoS especially for real-time applications, which are much sensitive to QoS fluctuation. Manuscript received December 17, 2011; approved for publication by Phone Lin, Division III Editor, July 4, 2012. This work was supported by the National Fundamental Research and Development Programs of China (2011CB707003) and the National Natural Science Foundation of China (61101083) and the Fundamental Research Funds for the Central Universities of China (2012ZM0021 and 2012ZZ0031), SCUT. The authors are with the School of Electronic and Information Engineering, South China University of Technology, P. R. China, email: {mian.guo123, huachaomao}@gmail.com, {shmjiang, eeqshguan}@scut.edu.cn. Digital Object Identifier 10.1109/JCN.2013.000011

The Internet engineering task force (IETF) has standardized two QoS models for the Internet: The per-flow integrated services (IntServ) and per-class differentiated services (DiffServ). IntServ can provide distinct per-hop services for every granted flow. However, it faces the scalability problem since routers in the IntServ network have to maintain state information for each granted flow. DiffServ tries to handle the scalability problem by using packet classification and packet marking schemes and aggregating flows belonging to the same class into one flow. In this case, the router only needs to process a limited number of classes, and each class is treated independently. Obviously, DiffServ may provide coarse QoS granularity for the flows aggregated into the same class. Now, most end-to-end QoS schemes can provide non-seamless QoS support in the wired-wireless integrated networks [4]–[8]. That is to say, these schemes provide end-to-end QoS support by partitioning the network into backbone and access segments, with DiffServ being used in the backbone segment while IntServ is used in the access segment. In this case, a QoS parameter mapping between the backbone and the access segments should be carried out at the network border. However, QoS parameter mapping may lead to other problems, such as extra mapping delays per packet, which may make such parameter mapping become a performance bottleneck. Recently, a per-packet differentiated queueing service (DQS) approach was proposed for wired networks initially and then, extended to wireless networks [9]–[11]. DQS tries to balance between implementation scalability and QoS granularity mainly by using two mechanisms. The first mechanism is that DQS explicitly requires each packet to carry its QoS requirements by itself in terms of end-to-end delay bound and packet loss preference for QoS scalability. The second mechanism is the enhancement of buffer admission control (BAC), by which each arriving packet is properly placed in the queue according to its QoS requirement and those of all the already queued packets. In this case, the output scheduler simply picks up the packet at the head of line for service. In this paper, we study how to use DQS to cost-efficiently support end-to-end QoSs in the network integrated of wired and wireless segments, particularly focusing on the provisioning of QoS adaptability to mobility and to bursty applications. To this end, we propose a scheme to determine the adaptive delay bound for every arriving packet at the router so that QoS adaptability can be provided by positioning the arriving packets in the queue according to their instantaneous delay bounds. Simulation studies show that, the proposed scheme can yield highest QoS-efficiency in comparison with IntServ and DiffServ. Further studies show that the scheme is also cost-efficient by providing similar QoS-efficiency while consuming less resources in comparison with the DQS implementation for wireless networks

c 2013 KICS 1229-2370/13/$10.00


62

JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 15, NO. 1, FEBRUARY 2013

proposed by [10]. The remainder of this paper is organized as follows. Section II provides an overview of the QoS provisioning of IntServ, DiffServ, and DQS. Section III discusses the proposed scheme to determine the delay bound for every arriving packet at any router. Section IV analyzes the performance of our proposal. Section V discusses the QoS adaptability of the proposed scheme in comparison with IntServ and DiffServ and Section VI further evaluates the QoS adaptability of the proposed scheme through simulation study. Finally, the paper is summarized in Section VII. II. QOS PROVISIONING OF INTSERV, DIFFSERV, AND DQS This section provides an overview of the QoS provisioning of IntServ, DiffServ, and DQS. The implementation of DQS is also discussed. A. Delay Guarantee of IntServ, DiffServ, and DQS For IntServ, if an end-to-end delay bound is given, the delay bound to be provided by every router is calculated during the resource reservation procedure along the path before the connection begins. Once the per-router delay bound is calculated and granted, it is stored at each router and would not be changed until a new resource reservation procedure is initiated. IntServ usually uses delay-based scheduler to support different delay bounds. The most general one is the earliest deadline first (EDF) scheduler [12]. DiffServ tries to reduce queueing delays by allocating a much higher service rate than the actual arrival rate to low-latency traffic. Different implementations can be adopted, such as class-based weighted fair queueing and weighted round robin scheduling algorithms, etc. Generally, DiffServ gives the highest service priority to the class required delay guarantee and adopts strict priority queueing for scheduling [13]. With DQS, if the delay bound of an arriving packet at the current router is given, the BAC only allow this packet to enter into the queue if its delay bound can be provided by the current router, and if its insertion into the queue will not under-provision QoS to the existing packets. All packets in the queue are positioned in an ascending order according to their allowable dwelling time at the router. If the BAC works well, all the packets inside the queue can be served within their delay bounds. B. Implementation of DQS Obviously, accurately estimating the delay bound of every arriving packet according to its end-to-end delay requirement is a key to providing QoS adaptability. The original delay bound n arriving packet at router i is calculated by i−1 for each D− j=1 d˜j − j=i+1 d˜j [9], where D refers to the end-toend delay bound of this packet and d˜j is the delay bound at router j. When j<i, d˜j is the actual delay that the packet has experienced at node j. However, it is difficult for node i to determine d˜j for j>i. In practice, an estimation algorithm is needed to approximate d˜j for j>i. In [10], a periodical broadcast mechanism is proposed to estimate the delay between nodes inside wireless networks. With

this proposal, a node periodically sends a probe packet containing a time stamp and its delay table to its neighbors. When a neighbor receives this packet, it firstly calculates the difference between these two time stamps, thus obtains an approximated one-way latency between the upstream node and itself. Then, it updates the delay table with the updated latency. When the period is due, it marks its own time stamp and its own delay table in a probe packet and sends it to others. Such operation repeats throughout the whole wireless network. Since such periodical broadcast consumes network resources while wireless resources are so precious and scarce that, this proposal is somewhat uneconomical, especially when part of nodes do not deliver traffic during a long period of time. III. A PROPOSAL In order to enable DQS to support QoS adaptability in the integrated networks, we have to accurately determine the maximum delay bound of a packet in the current router by estimating the delay bounds at downstream routers. To this end, we propose a scheme to calculate the dynamic delay bounds for the arriving packets at routers in the integrated network. According to the particular properties of various sub-networks, we propose three different algorithms to calculate delay bounds for routers in different locations. Since the delay bound estimation algorithms need delay information from other routers, in the following dynamic delay bound calculation, we assume the clocks are synchronous or time stamps from different sources can be mapped into a wall clock in the current router. A. Calculation of Dynamic Delay Bounds at a Border Router Scheduling packets according to their service priorities can avoid emergent packets overdue when congestion happens. Since long queueing delay is likely to happen at border routers when the network is congested, and meanwhile, delay bounds uniquely determine the service priority of an arriving packet in the DQS model. So, accurately estimating the delay bound for every arriving packet at the border router is very important to DQS. In this sense, we try to accurately estimate the delay from the border router to the destination node inside the wireless network and store it at the border router. The detailed algorithm as illustrated in Fig. 1 is described below. As shown in Fig. 1, when the wireless border router i receives a packet to be delivered into the wireless network, it checks its delay table to look for the delay between itself and the destination node. If the delay is not found, it uses (2) to calculate the delay bound that the arriving packet can tolerate at the current router. Otherwise, assume the delay Dˆn between the current border router and the destination node n stored in the delay table. Then, (1) is used to calculate the delay bound for the arriving packet. i−1 (1) d˜j − Dˆn di = D − j=1

where di represents the delay bound that the packet can tolerate at the border router i, D is the tolerable end-to-end delay of this packet, and d˜j represents the actual delay that this packet has experienced at router j.


GUO et al.: PROVISIONING OF QOS ADAPTABILITY IN WIRED-WIRELESS INTEGRATED... Border router

Begin

Packet arrives.

No

Table 1. Summary of the proposed delay bound calculation.

Destination node

Begin

Location of the current router In wired domain At wireless border Within wireless domain

Packet arrives .

The delay exists in Yes the delay table?

63

Use (3) to calculate the delay.

Delay bound di ˜ ˆ (D − i−1 j=1 dj − Dn )/hi i−1 ˜ D − j=1 dj − Dˆn i−1 ˜ (D − dj )/hi j=1

Update the local delay table. Use (4) to calculate the delay bound.

Use (2) to calculate the delay bound.

Encapsulate extra information to this packet. Queue this packet for forwarding .

Receive the delay update packet .

Transmit the packet.

Period is due ?

No

Yes Send a delay update packet to the border router. End

Transmit the delay update packet.

Update the delay table.

No

Time out for this destination ? Yes Delete the record . End

Fig. 1. The algorithm to determine the delay bound of an arriving packet at the border router.

Simultaneously, a time stamp and the border router’s address as extra information are encapsulated into the header of that packet for later delay calculation at the destination. Then, this packet is queued for forwarding. If the packet is successfully delivered to the destination, the destination calculates the delay Dˆn from the border router to itself by deducing the time stamp marked by the border router in the packet header. Then, it calculates the average delay Dˆn by (Dˆn +Dˆn (N − 1))/N , where Dˆn is the last average delay and N is the total number of packets currently received by the destination within a period, which is the time interval that the source receives a delay feedback from the destination. Then, the destination updates the local delay table with this new average delay. Note that, the delay table only stores the information on the delay from the border router to this node. When the above period is due, the destination sends back a delay update packet to the wireless border router, informing it of the latest delay from the border router to this destination. When the border router receives a delay update packet from the destination, it updates the delay table. The border router also uses a timer for each destination. If the delay information of some destination has not been updated for a period of time, we assume that this destination is inactive or unreachable, and delete its record accordingly to save the storage resource and avoid large deviation of delay bounds. B. Calculation of Dynamic Delay Bounds in Wireless Networks Packets from applications may travel over more than one hop inside a wireless domain. In order to efficiently utilize the precious wireless resources as well as obtaining a certain accuracy

of approximated delay bounds, we try to exploit some information available from the routing layer to acquire the number of remaining hops from the current node to the destination. We also assume that nodes along the remaining wireless path provision the same delay bounds for the packet. Thus, (2) is used to determine the delay bound of an arriving packet at any wireless node inside the domain. i−1 D − j=1 d˜j (2) di = hi where di is the delay bound that the packet can tolerate at wireless node i, D is the tolerable end-to-end delay of this packet, d˜j is the actual delay that this packet has experienced at node j, and hi is the number of remaining hops from the current wireless node to the destination along the path. C. Calculation of Dynamic Delay Bounds in Wired Networks Since routers inside wired networks are easier to provide stringent per-hop services than routers inside wireless networks, we assume that the wired routers downstream and the current router can guarantee the same per-hop delay for the arriving packet. Therefore, the delay bound is calculated by di =

D−

i−1 ˜ ˆ j=1 dj − Dn hi

(3)

where Dˆn is an estimated delay from the last hop of the wired network to the wireless destination, which is obtained by using the algorithm proposed in subsection III-A (see Fig. 1). If the arriving packet is forwarded to a wired destination, Dˆn is set to zero. hi is the number of remaining hops from the current router to the wireless border router along the path. D. Summary Our proposal to calculate the dynamic delay bound for an arriving packet at routers in the wired-wireless integrated networks is summarized in Table 1. IV. PERFORMANCE ANALYSIS Among various performance metrics, end-to-end delay and packet loss rate are common and important in measuring QoS provisioning. Generally, the packet loss rate is referred to as the ratio of the number of packets that are lost in the interior routers due to congestion to the total number of packets that are sent by the source. However, according to the service discipline of DQS, overdue packets may also be actively dropped by the BAC. Hence, the lost packets in DQS include packets lost by congestion and the dropped overdue packets.


64

JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 15, NO. 1, FEBRUARY 2013

A. Goodput Here, we also define goodput to measure the ratio of the number of successfully received packets that meet their end-to-end delay requirements to the total number of packets that are sent by the source. Assume every flow can be modeled with an exponentially bounded burstiness (EBB) [14]. Then, at router i along the path, the arrival traffic of flow j at any time interval [t0 , t0 +t] satisfies P {Aj [t0 , t0 + t] − ρj t > x} ≤ ae−bx , ∀t ≥ 0, ∀t0 ≥0, where ρj is the upper mean arrival rate and a and b as well is constant. Herein, A(t) represents the accumulative amount of traffic during time interval t. Although the available bandwidth may vary in wireless networks, we can assume a statistical service rate allocated to flow j is rij . Similarly, although the perhop and per-packet delay bounds of a flow in our proposal may be distinct, we define dji as the statistical delay bound of flow j at router i. In this case, the queue length distribution of this flow at router i is given by

comparing the analytical and simulation results requires complex statistical bandwidth and delay bounds. Due to the space limitation, it will be our future work. B. Resource Consumption

The motivation of our proposal is to balance between costefficiency and end-to-end QoS provisioning efficiency. Obviously, the algorithm proposed in subsection III-A to determine the delay from a wireless border router to a wireless destination also consumes resources. Therefore, here we mainly focus on the cost-efficiency in terms of overhead and storage consumption by the scheme proposed in subsection III-A (also see Fig. 1) in comparison with the proposal of Teng [10] for wireless networks. As shown in Fig. 1, the proposed scheme has two types of overhead. The first one is the extra overhead in the header of the arriving packets at the border router. It consists of a time F (x)=1−P (qij >x) = 1−P ( sup {Aj (t)−Aj (s)−rij (t−s)>x}) stamp and the border router’s address. The other type is from t−s≤dji the delay update packet, which contains the delay from the border router to the destination and the destination’s address. Asae−bx , x≤rij dji . (4) sume both our scheme and the one proposed by Teng use media ≈ 1− j 1−e−b(ri −ρj ) access control (MAC) address to represent the wireless node’s And the probability density function can be calculated by address. Since the overhead for a time stamp and a MAC adj f (x)=dF/dx≈abe−bx /(1−e−b(ri −ρj ) ). So the average queue dress are 2 bytes and 6 bytes, respectively [17], the data packet in the proposed scheme consumes extra overhead of 8 bytes per length can be obtained by packet. The payload of the delay update packet is 10 bytes per rij dji j j brij brij dji −ae (br d +1 − e ) packet. In the proposal of Teng [10], every node periodically i i xf (x)dx ≈ E(qij ) = . (5) j j j broadcasts a hello probe packet to their neighbors. The broadbr d br bρ be i i (e i − e j ) 0 cast information consists of a delay table containing all address Then, the average queueing delay given by the router is de- pair and the delays between these address pair as well as a time ¯ duced by dji =E(qij )/rij . Accordingly, the average end-to-end stamp of the source node. Obviously, in this scheme, there are ¯j 14 bytes per record in the delay table, thus the maximum length delay for the flow j is given by D¯j = H i=1 di , where H is the number of routers the flow should experience when it is trans- of a probe packet can reach (14n(n−1)+2) bytes, where n is the number of nodes in the wireless network. Accordingly, the mitted from the source to the destination. j j If we let the buffer size for the flow be larger than ri di , then whole process consumes much larger network resources as comthe lost packets only include dropped overdue packets. Simi- pared with our scheme. As to the storage consumption, in the proposal of Teng [10], lar to the stochastic delay bound probability calculation adopted each node is required to store the delay information for every in [15], in our proposal, for flow j at router i, the probability of node pair, which consumes lots of storage resources if the numthe actual delay exceeds the delay bound is bounded by ber of nodes is large. Differently, in our proposal, only the borj j der router and the active destination nodes are required to store ae−bri di j ˜j j j j . the delay information. Pi {di>di } = P {sup{Aj (t)−ri (t+di )} > 0}≤ j −b(r −ρ ) j i t≥0 1−e Accordingly, our proposal is more cost-efficient with less (6) According to [16], the packet loss rate can be approximated overhead and consuming less storage resources in comparison ˜ ˜ by Plrji {dji >dji }=αPij {dji >dji }, where α is a constant. There- with the scheme proposed by Teng. fore, similar to [15], the final packet loss rate of flow j is given C. Effects of Delay Bound Estimation Errors by H As mentioned in subsection II-A, the delay bound in the pro(1 − Plrji ). (7) Plrj = 1 − posal determines whether this packet can be inserted into the j=1 Since the buffer size is large enough, we can assume that all queue as well as the queueing position at current router. Generthe received packets in our proposal meet their end-to-end delay ally, mobility may affect the accuracy of the estimation. In fact, the estimation errors of delay bounds may cause packets to be requirements. Accordingly, goodput is approximated by dropped by the BAC or to experience longer delay than its endGj = 1 − Plrj . (8) to-end delay bound, resulting in G = 1 − Plr. The analysis model throws light on the factors that have imTherefore, the following formula is used to investigate the pact on the performance of our proposal. On the other hand, deviation of the estimation to flow j in a network that the buffer


GUO et al.: PROVISIONING OF QOS ADAPTABILITY IN WIRED-WIRELESS INTEGRATED...

65

size is large enough. Source 1

j =

˜ j 1 − Plr −1 G˜j

Router 1

Router 3

2 7

(9)

˜ j and G˜j are statistical packet loss rate and goodput, where Plr respectively. Accordingly, the mean deviation μ and variance σ of the proposal are obtained as μ = E( j ) and σ 2 = E( j − μ)2 (j ∈ [1, N ]), where N is the total number of input flows in the network. As μ approximates zero, the smaller σ is, the more accurate the estimation is. V. QOS ADAPTABILITY In this section, we further discuss QoS adaptability to mobility and to bursty applications. A. QoS Adaptability to Mobility QoS adaptability to mobility is required in wired-wireless integrated networks since mobility will cause QoS fluctuation and the wireless channel quality is very sensitive to mobility. Mobile nodes commonly move at various speeds. Generally, nodes move at low speed experience small fluctuation of channel quality while high speed can cause big fluctuation. For example, if a mobile node moves far away from the sender node or a relay node, the increasing speed can increase packet loss rate or prolong average delays. However, if it moves towards the above nodes, the opposite results will be yielded. Both IntServ and DiffServ adapt to mobility by dynamic QoS negotiation and resource reservation at the call level. Protocols such as resoure reservation protocol (RSVP) are used to carry out resource reservation for a granted flow in IntServ or for an aggregated class in DiffServ, respectively. However, while channel conditions vary frequently, it is impractical to frequently initiate resource reservation and QoS re-negotiation in ad hoc mobile networks. Particularly, with DiffServ, it is almost impossible to dynamically adjust the QoS for an aggregated flow following dynamic changes of an individual flow because of its aggregation mechanism. Our scheme can be easily adaptive to mobility due to its perpacket service granularity. According to the service discipline of DQS, the per-hop service to every arriving packet is determined by the packet’s position in the queue. This position is determined instantly according to the arriving packet’s tolerable QoS at the current router, such as delay bound. According to our scheme, this delay bound is dynamically adjusted according to the delay that this packet has experienced at the upstream nodes and the delay that the packet will experience at the remaining journeys. In this sense, the prolonged delay caused by mobility can be compensated by dynamically adjusting the packet’s dwelling time at routers along the path. For example, the terminal movement affects the channel quality, resulting in delay fluctuation over this segment of the path. With our scheme, the packet to be delivered to this terminal is likely to be served ahead at upstream nodes so that the prolonged delay caused by the movement is compensated. Consequently, QoS adaptability to mobility is yielded.

Source 2 ... Source n

Router 2 Backbone

13

8

Border router 17 1 16

3

4 11 5

The ad hoc network

Traffic sources

Fig. 2. The wired-wireless integrated network model for the simulation.

B. QoS Adaptability to Applications Since real-time applications are much sensitive to QoS fluctuation, the QoS provisioning should be adaptive to real-time applications in order to provide a stable QoS support. IntServ provides QoS for real-time applications by reserving adequate bandwidth and buffer resource for every granted flow, e.g., reserving service rate larger or equal to the peak arrival rate. However, if the traffic load of the granted application is low, the reserved resource will be wasted. This is because that less flows can be granted to the network if large bandwidth resources are reserved to bursty low traffic load applications. It is particularly uneconomical to do so in the wireless network since wireless resources are scarce and precious. Furthermore, it is difficult to reserve stable resources in wireless networks due to the dynamic nature of wireless channel capacity. Similar to IntServ, DiffServ provides QoS for bursty traffic by provisioning much larger bandwidth than the mean arrival rate of the bursty traffic, namely capacity over-provisioning. However, the lack of a per-flow granularity makes it possible for the applications to congest each other. For example, in a bursty period of a flow, all flows belonging to that class suffer QoS deterioration [18]. Differently, our scheme adapts to bursty applications by allocating an adaptive delay bound for every arriving packet at routers along the path. Every packet’s delay bound at a router is determined by its end-to-end delay requirement, the actual delay at upstream routers as well as the delay bound at downstream nodes. During the period of a burst, a number of packets may experience long queueing delays at a router while other packets prolong their delays at other routers. However, their end-to-end delays may still be bounded since the sum of their per-router delays along the path would not exceed the end-to-end delay bound according to the adaptive delay bound calculation algorithm proposed in Section III. VI. SIMULATION EVALUATION This section further evaluates the QoS adaptability of the proposed scheme by investigating the QoS efficiency for real-time applications through simulation in NS-2 in comparison with IntServ and DiffServ, particularly in terms of the adaptability to mobility and to bursty applications discussed in Section V. Simulation studies are also conducted to evaluate the efficiency of our algorithm in wireless networks in comparison with the algorithm proposed by Teng [10] to extend the discussion of subsection II-B and subsection IV-B. Finally, the accuracy of the proposal is investigated.


66

JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 15, NO. 1, FEBRUARY 2013

The simulated wired-wireless network topology is shown in Fig. 2, which consists of a wired network and a wireless ad hoc network based on the 802.11b standard. The wired network includes multiple traffic sources and three fixed backbone routers. The ad hoc network has 20 nodes randomly distributed in a grid of 800 m × 800 m with a transmission rate of 2 Mbps. A wireless border router with a capacity of 5.5 Mbps interconnects the wired and wireless networks. Since the capacity of the ad hoc network is limited here, we set the bandwidth of the wired routers to 10 Mbps. Reference [19] points √ out that the optimal capacity per-node approximates Θ(C/ n) for a destination non-vanishing far away, where C is per-node capacity and n is the number of nodes in a disk of unit area. If both mobility and delay requirements are considered, the per-node capacity may be far lower. Therefore, we constrain our simulated traffic to a reasonable load to keep throughput as high as possible while given good QoS performance in static optimal situation. Then, we investigate the performance in various situations. In general, real-time applications are much sensitive to delay while some packet losses may not obviously degrade the decoding quality. Accordingly, we do not retransmit lost packets in this paper. However, we may consider the effect of retransmition in the future work. We use the following performance metrics in the discussion. • End-to-end delay: The time taken by a successfully received packet traveling from source to destination. • Goodput: The ratio of the number of successfully received packets that satisfied their end-to-end delay requirements to the total number of packets that are sent by the source. • Packet loss rate: The ratio of the number of packets that are lost due to congestion or delay bound violation to the total number of packets that are sent by the source. • Overhead: The extra per node throughput that are transmitted between routers for delay calculation. • Storage: The average number of records that per node should store for delay calculation. A. QoS Adaptability versus Mobility In this case, multiple voice over Internet protocol (VoIP) flows are sent from the wired traffic sources to the nodes randomly located in the ad hoc network. The data rate of each VoIP flow is set to 80 kbps according to the G.711 standard, while its endto-end delay bound is set to 150 ms. Two scenarios are simulated. The first one is that 10 VoIP flows are sent from the wired sources to 10 wireless destination nodes with only one destination moving. The other scenario is that, all destinations randomly move at the maximum speed of 100 m/s and the number of VoIP flows is increased. In both scenarios, all nodes are randomly distributed around the wireless border router at the beginning, and then the mobile nodes move randomly within the grid during the simulation. Here, IntServ conducts resource reservation and adopts EDF to schedule flows throughout the wired-wireless network. According to [12], the reserved rate and bucket depth for each flow per router is set to 80 kbps and 1000 bytes, respectively. During the simulation, if the connection setup is unsuccessful or the connection is interrupted due to temporary channel quality dete-

rioration, the resource reservation procedure will be re-initiated after a period of time until it is successful. For DiffServ, since every flow has a distinct address pair, we constrain the arrival rate of each flow to 80 kbps by using the token bucket at the network boundary according to [20]. All flows are aggregated to the highest per-hop behavior (PHB) group and obtain the same perhop service according to their similar traffic characteristics and end-to-end delay requirements. At the wireless border router, all VoIP flows are set to the highest service priority. A strict priority queueing is used to schedule them inside the wireless network. Non-compliant packets will be queued to the lowest priority queue. With the proposed scheme based on DQS, we mark 150 ms as the end-to-end delay bound in the header of each VoIP packet when it is generated. If a packet is rejected by the BAC from queueing it in the QoS guaranteed queue, the packet is queued in the best-effort queue. If a packet is overdue in an intermediate node, it is dropped immediately. As discussed in subection IV-A, the buffer size is constrained by the delay bound and the service rate. If the buffer size is larger N than i=1 ri di , where ri is the service rate for flow i, di is the delay bound of flow i, and N is the number of flows, then the lost packets only consist of dropped overdue packets in our proposal. As to both IntServ and DiffServ, if the buffer size is larger than that threshold, some packets will stay a long time in the queue and thus, will become overdue when they are served. However, if the buffer length is shorter than that threshold, some normal packets will be discarded. Therefore, the buffer sizes of the above three schemes are all set to 150 packets, which is a small larger than N i=1 ri di , such that the normal packets will not be discarded. We first look at the QoS performance of the flow to the mobile node for scenario 1 as depicted in Fig. 3. As shown in Fig. 3, when the mobile node moves at a low speed, all the three QoS schemes can guarantee the QoS of all VoIP flows; however, as the speed increases, their performance declines obviously. The reason is that, during a node’s movement, the node may suffer serious interference and/or travel to a blind zone, in which the node is isolated and cannot communicate with other nodes. Since the node (note that here only one node is moving) is located around the wireless border router at the beginning of simulation, the channel quality may suffer slightly when the node moves at a low speed. However, as the speed increases, the interference between it and the other active nodes and the probability for it to walk into blind zones become high, and accordingly, the channel quality fluctuates more frequently as the speed increases, resulting in performance deterioration. Overall, the end-to-end delay given by the proposed scheme is obviously lower than those given by the other two schemes as shown in Fig. 3(a). The goodput given by the proposed scheme is also higher than those given by the other two as shown in Fig. 3(b). As for the performance of all flows including flows to the mobile destination and to the static destinations as depicted in Fig. 4, the simulation results show that, although both our proposal and IntServ schedule packets according to their delay bounds, they yield different QoS performance. The main reason is their distinct mechanisms to obtain the delay bound for each packet. With our proposal, the per-hop service priority of every packet is determined by each packet’s instantaneous delay


GUO et al.: PROVISIONING OF QOS ADAPTABILITY IN WIRED-WIRELESS INTEGRATED...

1

0.3

0.25

0.25 0.2

0.2

0.15

Delay bound

Packet loss rate

0.9

Goodput

End−to−end delay (s)

67

0.8

0.7

2

5

10

20

60

80

0.6

100

0.1

0.05

0.1

0.05

0.15

2

5

10

Speed (m/s)

20

Speed (m/s)

(a)

60

80

0

100

IntServ DiffServ The proposal 2

5

10

(b)

20

Speed (m/s)

60

80

100

(c)

Fig. 3. The performance of the flow to mobile node for scenario 1: (a) End-to-end delay, (b) goodput, and (c) packet loss rate. There are total 10 flows, the end-to-end delay is 0.15 s, and only one destination moves.

0.16

0.2

0.9

0.15

Packet loss rate

Delay bound

0.14 0.12

Goodput

End−to−end delay (s)

1

0.1

0.8

0.08

0.7

0.1

0.05

IntServ DiffServ The proposal

0.06 0.04

2

5

10

20

Speed (m/s)

60

80

100

0.6

2

5

10

20

Speed (m/s)

(a)

(b)

60

80

100

0

2

5

10

20

Speed (m/s)

60

80

100

(c)

Fig. 4. The performance of all flows for scenario 1: (a) End-to-end delay, (b) goodput, and (c) packet loss rate. There are total 10 flows, the end-to-end delay is 0.15 s, and only one destination moves.

bound at the router. This delay bound is dynamically determined by considering the effect of mobility. Generally speaking, the per-hop delay bound of a packet toward a mobile node is probably smaller than those toward a static node. Therefore, they may have higher service priorities at interior routers. That is to say, upstream routers may shorten their queueing delay. Accordingly, before the destination node becomes unreachable, most packets can still be delivered to the mobile destination within their delay bounds. However, with IntServ, since the delay bound of every flow at a router is determined at the call level, the service priority of a flow toward a mobile node cannot be adjusted dynamically according to the instantaneously fluctuating channel quality. Consequently, the efficiency of QoS support in delivering packets to the mobile destination with IntServ is lower in comparison with the proposed scheme, resulting in a worse QoS performance as shown in Fig. 3. The movement of mobile nodes also leads to the channel quality fluctuations for other nodes because nodes in the ad hoc network have to compete for wireless resources with neighbors. Since IntServ is unable to quickly adapt to such instantaneous channel quality fluctuation as mentioned earlier, it also provides a worse QoS performance for static flows in comparison with our proposal, resulting in worse QoS performance as illustrated in Fig. 4. Due to the coarse granularity and static QoS mechanism, it is not surprise to see the worst QoS performance given by DiffServ in terms of end-to-end delay and goodput.

Note that, the packet loss rate given by DiffServ increases faster than those given by the other schemes as the speed increases as illustrated in Figs. 3(c) and 4(c). The reason is that all flows have the same traffic characteristic and QoS requirements as well as path situations, so they are serviced in a first in first out (FIFO) manner with DiffServ. Accordingly, as the speed increases, the prolonged transmission delay from the wireless border router to the mobile node increases the queueing delay of packets to other nodes and may cause congestion, as discussed in subsection V-B. However, since both our proposal and IntServ use delay-based scheduling, the prolonged transmission delay from the border router to the mobile node has small effect on the packets to other nodes. The simulation results for scenario 2 further reflect their different QoS adaptabilities to mobility. As shown in Fig. 5, the QoS differences between the three QoS schemes become more and more obvious as the number of flows increases. Since both our proposal and IntServ use delay-based packet scheduling, they can bound the end-to-end delay approximately to 150 ms. However, due to their different adaptabilities to mobility, the goodput given by our proposal obviously higher than that given by IntServ as the number of flows increases as shown in Fig. 5(b). Since DiffServ is least adaptive to mobility, it is not surprise to find that the end-to-end delay given by DiffServ increases fast and its goodput decreases nearly to zero when the number of flows reaches 12.


68

JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 15, NO. 1, FEBRUARY 2013

1

0.7

0.8

0.6

0.4 0.3 0.2

Delay bound

Packet loss rate

0.5

Goodput

End−to−end delay (s)

0.8 0.7

0.6

0.6

0.4

0.4 0.3 0.2

0.2

0.1 0

0.5

IntServ DiffServ The proposal

0.1

0 1

5

9

13

17

20

Number of VoIP flows

1

5

9

13

17

Number of VoIP flows

(a)

0

20

2

4

6

8

10

12

14

Number of VoIP flows

(b)

16

18

20

(c)

Fig. 5. Simulated results for scenario 2: (a) End-to-end delay, (b) goodput, and (c) packet loss rate. All nodes move, the maximum speed is 100 m/s, and the end-to-end delay is 0.15 s.

B. QoS Adaptability versus Applications 450

Arrival rate IntServ DiffServ The proposal

400 350

Rate (kbps)

Due to the network capacity constraint, in scenario 3, we send totally 6 bursty video flows from the sources in the wired network to the destinations located in the ad hoc network. The mean rate of each flow is set to 320 kbps and the peak rate is about 460 kbps. The maximum packet size is 600 bytes and the end-to-end delay bound of each flow is 200 ms. In order to investigate QoS adaptability to bursty applications, here the nodes inside the ad hoc network are set to be stationary. According to [12], we reserve 460 kbps for every flow per router along the path and use EDF to schedule all packets for IntServ. The per-flow delay bound at a router is calculated by Ctot /R+Dtot , where Ctot is set to the maximum packet size of the flow, R is the reserved rate for the flow, and Dtot is set to M T U/c, here c is the constant service rate of the router and M T U is the maximum transmission unit of the network. DiffServ constrains the arrival rate to 460 kbps for each destination at the edge of the DiffServ domain. If the amount of instant traffic exceeds the negotiated rate, the extra traffic will be degraded or dropped. All flows are aggregated to the highest class at the edge and acquire the same per-hop service according to their similar traffic characteristic and QoS requirements. At the wireless border router, these bursty flows are mapped to the highest service priority and a strict priority queueing scheme is used to schedule them in the wireless network. Our proposal just marks 200 ms as the end-to-end delay bound for every packet when it is generated. Similar to the reasons mentioned in subsection VI-A, the buffer sizes of the three schemes are set to 500 packets in this simulation, which is enough for the burst by considering the delay requirements and the channel capacity constraints. Other implementation issues are similar to those mentioned in subsection VI-A. When multiple bursty real-time flows simultaneously feed into the network, the accumulative queue length of the bottleneck router will increase quickly. Accordingly, many packets may experience especially long queueing delays. Since all the bursty flows in the simulation have similar traffic characteristic, paths and end-to-end QoS requirements, they acquire identical per-hop service in both IntServ and DiffServ. Particularly with DiffServ, the lack of per-flow granularity makes heavy conges-

300 250 200 150 100

0

50

100

150

200

250

300

350

400

Simulation time (s) Fig. 6. Rate of bursty real-time flow 3 for scenario 3.

tion during the period of bursts, resulting in much longer queueing delays as shown in Table 2. Accordingly, DiffServ yields the longest average end-to-end delay and lowest goodput as shown in Table 2. Differently, with our proposal, although a packet may experience a long queueing delay at the bottleneck router, its perhop queueing delay can be shortened at other routers according to the per-packet adaptive delay bound mechanism proposed in Section III. Therefore, more packets can meet their end-to-end delay requirements in comparison with both IntServ and DiffServ. Accordingly, the QoS performance given by our proposal is the best with the shortest average end-to-end delay and highest goodput as shown in Table 2. The adaptability of our scheme to applications can be further demonstrated in Fig. 6, where the effective service rate given by our proposal is quite closer to the arrival rate while that given by DiffServ deviates much during the period of bursts. Therefore, it is not surprise to see that the goodput given by our proposal is stable while the goodput given by DiffServ fluctuates much as illustrated in Table 2. Note that, although the average goodput given by our scheme is higher than that given by IntServ, the goodput of some flows given by IntServ may be better than that given by our scheme as shown in Table 2. The reason is that, the per-flow QoS pro-


GUO et al.: PROVISIONING OF QOS ADAPTABILITY IN WIRED-WIRELESS INTEGRATED...

69

3

10

1 VoIP the proposal VoIP Teng Video the proposal Video Teng

1

0.6

0.4

0.2

10

Storage (record/node)

Overhead (kbps/node)

Goodput

0.8

0

10

The proposal Teng 0

2

4

6

8

10

12

14

16

18

2

Number of flows

4

6

8

10

12

Number of flows

(a)

14

16

18

2

10

1

10

0

10

The proposal Teng 2

(b)

4

6

8

10

12

Number of flows

14

16

18

(c)

Fig. 7. Cost-efficiency comparison between our proposal and Teng’s (the number of VoIP flows against the number of video flows is 1:1) (scenario 4): (a) Goodput, (b) overhead, and (c) storage consumption. Table 2. Performance for bursty real-time flows (scenario 3).

Flow 1 2 3 4 5 6 Mean

End-to-end delay (s) PropoIntServ DiffServ sal 0.179 0.191 0.143 0.179 0.192 0.124 0.181 0.192 0.124 0.181 0.193 0.124 0.180 0.192 0.125 0.179 0.193 0.124 0.180 0.192 0.128

Goodput IntServ

DiffServ

0.979 0.973 0.886 0.908 0.953 0.977 0.946

0.721 0.702 0.684 0.664 0.642 0.623 0.673

visioning mechanism may provide better service for a flow toward a quite stable destination but providing the worse services for other flows toward unstable destinations when they have the same traffic characteristic, end-to-end QoS requirements, and travel along the same path. Differently, our scheme tries to provide fair goodput for these flows by scheduling every packet according to its QoS requirement. Thus, as shown in Table 2, the goodput of every flow given by our proposal is quite closer to that of the average one. C. Cost-Efficiency Our scheme determines the delay bound at nodes in the wireless network in a way quite different from the proposal of Teng [10], which was discussed earlier in subsection II-B. As discussed in subsection IV-B, our scheme is cost-efficient by consuming lower network resources. In the following, we further validate the cost-efficiency of our proposal by comparing with the proposal of Teng in wireless networks. To this end, we send VoIP and bursty video traffic from the wireless border router to the nodes inside the ad hoc network in scenario 4. The traffic characteristic and QoS requirements of both types of applications are similar to those used in subsections VI-A and VI-B, respectively. During the simulation, every node moves at speeds ranging from 2 m/s to 100 m/s and each flow is transmitted to a distinct wireless node. We investigate the efficiency by increasing the number of flows. As shown in Fig. 7(a), the goodput of both applications given by our scheme is quite closer to that given by Teng. However, the overhead and storage consumption are far lower as shown in

Proposal 0.969 0.965 0.967 0.961 0.960 0.967 0.965

Packet loss rate PropoIntServ DiffServ sal 0.008 0.015 0.029 0.011 0.021 0.032 0.027 0.027 0.031 0.038 0.036 0.035 0.026 0.048 0.036 0.010 0.064 0.032 0.020 0.035 0.031

Figs. 7(b) and 7(c). The reason for the similar goodput is that, the queue length at a wireless node inside the ad hoc network is far shorter than that at the border router since the traffic load in an interior node is far lower than that in a border router. Therefore, scheduling packets according to their QoS requirements at the border router is much important for QoS provisioning than at an interior wireless node. Accordingly, both schemes acquire similar goodput as shown in Fig. 7(a) since they have considered the delay from the border router to the destination node. As discussed in subsection IV-B, our proposal transmits delay related information between border routers and active destination nodes. This information is encapsulated in the header of the arriving packets at the border router. So the overhead increases with the number of flows as shown in Fig. 7(b). On the contrary, the overhead with the proposal of Teng is proportional to the number of nodes. So the overhead is a constant if the number of nodes is fixed. However, the overhead consumption by the proposal of Teng is always higher than that by our proposal as shown in Fig. 7(b) because every delay probe packet used in the proposal of Teng has to carry delays between all address pairs. Regarding the storage consumption, similar to the overhead consumption, both the border router and the active destination nodes only record delays between these nodes with our proposal while all nodes have to record delays for all address pairs with the proposal of Teng. Thus, the storage consumption given by our proposal is far lower than that given by the proposal of Teng as shown in Fig. 7(c). Therefore, our proposal is quite efficient by yielding the lower resource consumption while acquiring similar QoS as compared with the proposal of Teng.


70

JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL. 15, NO. 1, FEBRUARY 2013

Table 3. Deviation of the estimation.

Scenario Mean (μ) Variance (σ 2 )

1 0.0085 1.9e-05

2 0.0138 9.5e-05

3 0.0030 1.3e-06

4 0.0230 5.0e-04

D. Estimation Errors As discussed in subsection IV-C, the delay bound is determined approximately. The straight effect of estimation errors reflects in G =1−Plr. Table 3 lists the mean deviation and variance of the delay bound estimation errors in various simulation scenarios discussed in the paper. As shown in Table 3, the mean and variance given by scenarios 2 and 4 are obviously higher than those given by the other two scenarios, which shows that the deviation increases as more nodes are moving. However, the deviation is still within an acceptable range. VII. CONCLUSIONS QoS adaptability support to mobility and traffic burstiness is important since these features are common in wired-wireless integrated networks with dynamic traffic loads and wireless network capacity. This paper proposed a scheme to dynamically determine the delay bounds at the packet level based on DQS to support QoS adaptability in such kind of networks. Simulation studies show that, the proposed scheme can provide better QoS adaptability in this kind of network in terms of better QoS performance in comparison with both IntServ and DiffServ. The proposal is also more cost-efficient than another DQS-based scheme proposed by Teng with similar QoS performance but consuming less network resources. ACKNOWLEDGMENT We thank the editor and the reviewers for their constructive comments, which have improved the quality of this paper. REFERENCES [1] [2] [3] [4]

[5] [6] [7] [8] [9]

4G americas, 4G mobile broadband evolution: 3GPP release 10 and beyond-HSPA+, SAE/LTE and LTE-Advanced. [Online]. Available: http: //www.4gamericas.org/ A. Sanchez-Esguevillas, B. Carro-Martinez, and V. Poosala, “Future convergent telecommunications services: Creation, context, P2P, QoS, and charging,” IEEE Commun. Mag., vol. 49, no. 1, pp. 58–59, Jan. 2011. P. TalebiFard, T. Wong, and V. C. M. Leung, “Access and service convergence over the mobile Internet - A survey,” Comput. Netw., vol. 54, no. 4, pp. 545–557, Mar. 2010. M. A. Callejo-Rodrigitez, J. Enriquez-Gabeiras, W. Burakowski, A. Beben, J. Sliwinski, O. Dugeon, E. Mingozzi, G. Stea, M. Diaz, and L. Baresse, “EuQoS: End-to-end QoS over heterogeneous networks,” in Proc. K-INGN, 2008, pp. 177–184. S. Senkindu and H. A. Chan, “Enabling end-to-end quality of service in a WLAN-wired network,” in Proc. IEEE Int. Conf. Mil. Commun., 2008, pp. 1–7. P. Stuckmann and R. Zimmermann, “European research on future Internet design,” IEEE Wireless Commun., vol. 16, no. 5, pp. 14–22, 2009. X. Dong, J. Wang, Y. Zhang, M. Song, and R. Feng, “End-to-end QoS provisioning in future cognitive heterogeneous networks,” in Proc. ICCPA, 2009, pp. 425–429. R. Good, D. Waiting, and N. Ventura, Quality of Service Provisioning in the IP Multimedia Subsystem. IGI Global, 2010, pp. 443–463. S. Jiang, “Granular differentiated queueing services for QoS: Structure and cost model,” Comput. Commun. Rev., vol. 35, no. 2, pp. 13–22, 2005.

[10] X. Teng, S. Jiang, G. Wei, and G. Liu, “A cross-layer implementation of differentiated queueing service (DQS) for wireless mesh networks,” in Proc. IEEE VTC Spring, 2008, pp. 2233–2237. [11] S. Jiang, Differentiated Queueing Service (DQS) for End-to-End QoS Provisioning: An Evaluation from Per-Flow, Per-Class to Per-Packet. Nova Science: NY, 2011, pp. 13–28. [12] J. Boudec and P. Thiran, Network Calculus - A Theory of Deterministic Queuing Systems for the Internet. Springer, 2004. [13] T. Szigeti and C. Hattingh, End-to-End QoS Network Design: Quality of Service in LANs, WANs, and VPNs. Cisco Press, 2004, p. 768. [14] O. Yaron and M. Sidi, “Performance and stability of communication networks via robust exponential bounds,” IEEE/ACM Trans. Netw., vol. 1, no. 3, pp. 372–385, June 1993. [15] S. Jiang, Future Wireless and Optical Networks: Networking Modes and Cross-Layer Design. Springer, 2012. [16] H. S. Kim and N. B. Shroff, “Loss probability calculations and asymptotic analysis for finite buffer multiplexers,” IEEE/ACM Trans. Netw., vol. 9, no. 6, pp. 755–768, Dec. 2001. [17] ANSI/IEEE Std 802.11:1999(E) Part 11: Wireless LAN Medium Layer (PHY) Specifications, ANSI/IEEE Standard 802.11, 1999. [18] T. Szigeti, End-to-End QoS Network Design. Cisco Press, 2004. [19] M. Grossglauser and D. Tse, “Mobility increases the capacity of ad-hoc wireless networks,” in Proc. INFOCOM, 2001, pp. 1360–1369. [20] S. Vegesna, IP Quality of Service. Cisco Press, 2001. Mian Guo (S’11) received her B.Eng. degree and M.S. degree in South China University of Technology, China, in 2002 and 2007, respectively. She is currently a Ph.D. student with the School of Electronic and Information Engineering, South China University of Technology. From 2002 to 2010, she worked in China Unicom Maoming Branch, China. Her major research area is quality of service provisioning for wired-wireless integrated networks.

Shengming Jiang (A’96-M’00-S’07) received his B.Eng. degree from Shanghai Maritime Institute, China, in 1988, DEA and Dr. degrees, respectively from University of Paris VI and University of Versailles Saint-Quentin-En-Yvelines, France, in 1992 and 1995. Currently, he is a Professor in the School of Electronic and Information Engineering, South China University of Technology. His major research area is communication networks.

Quansheng Guan (S09-M11) received the B.Eng. degree in Electronic Engineering from Nanjing University of Post and Telecommunications, China, in 2006 and the Ph.D. degree from South China University of Technology, China, in 2011. He is currently faculty member with the School of Electronic and Information Engineering, South China University of Technology. His research areas include topology control, routing and cooperative communications for mobile ad hoc networks, and cognitive networks.

Huachao Mao received the B.Eng. degree from South China University of Technnology, China, in 2010. He is currently a postgraduate with the School of Electronic and Information Engineering, South China University of Technology. His research areas include cross-layer architecture design for multi-hop wireless networks, especially for congestion control in such kind of networks.


Turn static files into dynamic content formats.

Create a flipbook
Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.