A Cross-Layer Implementation of Differentiated Queueing Service ...

2 downloads 206 Views 137KB Size Report
Index Terms—Wireless Mesh Networks (WMNs), Scheduling algorithms, Cross-layer implementation, Differentiated queueing service (DQS). I. INTRODUCTION.
A Cross-Layer Implementation of Differentiated Queueing Service (DQS) for Wireless Mesh Networks Xuefang Teng, Shengming Jiang (Senior member, IEEE), Gang Wei (Member, IEEE) and Guikai Liu School of Electronic and Information Engineering South China University of Technology Guangzhou, China Email: {tengxuefang,gkliu698}@gmail.com, {shmjiang,eegwei}@scut.edu.cn

Abstract—Provisioning of more granular quality of services (QoS) for different applications in wireless mesh networks (WMNs) is a challenge issue for scheduling algorithms. Most traditional scheduling algorithms try to provide per-flow or per-class QoS almost without consideration of the end-to-end situation for applications such as path lengths. Actually, for the same type of applications, difficulties for a node to provide end-to-end QoS largely depend on path lengths. Differentiated queueing service (DQS) is a new scheduling algorithm providing more granular QoS in wired networks by taking into account the above issue. In this paper, DQS is improved such that it can be used in WMNs. Meanwhile we also discuss a cross layer implementation of DQS. The earliest deadline first (EDF) scheme is used in comparison with our proposed algorithm since EDF is similar to DQS to some extent. The simulation studies show that the performance of DQS is better than EDF in terms of average goodput and end-to-end delay. Index Terms—Wireless Mesh Networks (WMNs), Scheduling algorithms, Cross-layer implementation, Differentiated queueing service (DQS).

I. I NTRODUCTION The wireless mesh networks (WMNs) [7], [10], [11] is an emerging technology for next-generation wireless networks to provide the last-mile Internet access for users. In future networks, there will be more and more types of applications requiring more granular quality of service (QoS) guarantees. As a new network technology, WMNs has to be able to support different levels of QoS rather than a single, best-effort service or gross QoS to support various applications. Regarding the provisioning of different levels of services in networks, paper [12] pointed out two paths to achieve this, namely control path and data path. Data path mechanisms try to provide different services through classifying and mapping user packets to the intended service class as well as controlling the amount of network resources that each service class can consume. Control path mechanisms are used to allow the users and the network to agree on service definitions, identify which users are entitled to the given service and let This research was supported partially by The National Hi-Tech Research and Development Program of China (863) under Grant No. 2007AA01Z210. Jiang currently is also with the Faculty of Advanced Technology, University of Glamorgan, UK.

978-1-4244-1645-5/08/$25.00 ©2008 IEEE

the network appropriately allocate resources to each service. Data path mechanisms are critical to the provisioning of QoS for applications since they define what action will be taken on the data packet so that some level of service can be satisfied. Two important resources in network are buffer and bandwidth. Buffer management and scheduling algorithms are the corresponding mechanisms to manage these resources. Buffer management decides which packet can be stored in buffer, while scheduling algorithms decide which packet can be transmitted next, and they are often used jointly. There are many scheduling algorithms having been proposed for wireless networks. The typical ones include CSDPS+CBQ, the Idealized Wireless Fair Queuing (IWFQ) [6], [13] and the Earliest Deadline First (EDF) [9] etc. In CSDPS+CBQ which combines the channel state dependent packet scheduling (CSDPS) and class-based queuing (CBQ) [3], traffic flows are grouped into classes, and each class is committed with a certain amount of bandwidth. This algorithm enables fair sharing of the wireless channel while trying to maintain high throughput, but it does not resolve other QoS issues such as delay bound and loss rate guarantee [15]. IWFQ is a classic wireless fair scheduling algorithm which aims to handle location-dependent error bursts. EDF is a dynamic priority scheduling algorithm which schedules packets in the order of their arriving times plus a static delay. One of the main merits of EDF is that it allows the separations of delay and throughput guarantees for a flow [12]. The above scheduling algorithms have a common characteristic, i.e., the service granularity is per-flow or per-class based so that the scheduler must know exactly per-class or perflow QoS requirements accordingly. Such per-flow treatment with storing voluminous information in network units is the major cause of the scalability problem and will complicate the implementation. To avoid this problem, [14] proposed a new scheduling algorithm called Differentiated Queuing Service (DQS) to provide end-to-end QoS cost-effectively according to the QoS requirements and end-to-end situation of applications such as path lengths. However, DQS was originally proposed for wired networks. In this paper, we will investigate how to improve it for wireless mesh networks by using a cross-layer

2233

Control packet queue (FIFO)

implementation, which tires to use the information available for the routing operation to support the DQS scheduling algorithm. The reminder of this paper is organized as follows. Section II describes the principle of DQS and a cross-layer implementation for wireless mesh networks in detail. Section III discusses simulation results by comparing DQS with EDF. Finally, the paper is summarized in Section IV. II. DQS FOR WIRELESS MESH NETWORKS The original DQS [14] was proposed to provide more granular QoS in wired networks with the following assumptions. Any application should have a maximum end-to-end delay that can be defined exactly. Therefore this delay is used as the lifetime of the associated packets. When a node receives such a packet, it will check whether the packet’s lifetime has expired. If so, the node should drop the packet immediately or put it into a best-effort service queue. The major advantage of DQS is its strong capability to support granular end-toend QoS more cost-effectively in terms of resource utilization. Therefore, it is interesting to apply DQS for wireless mesh networks to support end-to-end QoS. A. Principle and implementation issues of DQS The following notation is used in the discussion. • a: packet’s arrival time. • e: latest time for the departure of a packet from a node subject to its end-to-end delay. • D: packet’s maximum end-to-end delay. • T: packet’s maximum remaining lifetime upon its arrival at a node. Subscript ’i’ is associated with these parameters to indicate those for node i. To simplify discussion, the propagation delay is not addressed here since it is a constant for a given path. Therefore, for a path consisting of n nodes, we can have the following relationships between the above parameters for a packet arriving at node i. Ti = D −

i−1 

dj = Ti−1 − di−1

(1)

j=1

with T1 = D, where di is the actual delay that the packet experiences at node i. The effective packet’s maximum delay (which is defined as the interval between the arrival of the first bit and the departure of the last bit from the node for the same packet) allowed at node subject to its end-to-end delay di , is given by n  (2) dj . di = Ti − j=i+1

If the upstream nodes have bounded the delay as promised, then (3) ei = ai + di , which is used to differentiate packet queueing in the buffer. From above description, we can find that this algorithm schedules a packet according to its end-to-end delay situation rather

Packets

Out packets Classifier

Selector Data packet queue (DQS)

Fig. 1.

Structure of the scheduler in a node

than its class or flow. This means that two packets belonging to the same class or flow may be still treated differently if they travel along different end-to-end paths. Therefore, this algorithm can provide more granular end-to-end QoS support. B. A DBF-based Implementation for WMNs From (2), dj (for j > i) should be available for node i to determine ei with (3). However, this is difficult since dj will be only available in the future rather than upon the packet’s arrival at node i. Therefore, one of the implementation issues for DQS is how for node i to estimate dj for the remaining journey of a packet. This will affect the end-to-end delay guarantee and the network resource utilization. Another implementation issue is how to queue control packets since these packets usually do not have explicit end-to-end delay requirements so that they cannot be inserted into the same queue as that for data packets. However, control packets usually require privileged services. In the following section, we discuss how to resolve these problems in wireless mesh networks. 1) Estimation of delay over the remaining journey: Although it is difficult to get the exact delay for the remaining journey of a packet, this delay can be estimated according to the current path situation as discussed below. A node can send a probe packet with a time stamp on it to its neighbors. When a neighbor receives the packet, it marks its own time stamp over it and sends the packet to others. The difference between these two time stamps is just a one-way delay. So the node can learn the delay of a packet received from one of its neighbors. From this point of view, every node can maintain a delay table which records the delay between any nodes in networks through the above broadcasting operation. In fact, this is the principle of the famous routing protocol, Distributed Bellman-Ford (DBF) [4], which is a proactive table-driven protocol. This protocol consists of the following steps. A node calculates the delay between it and all other nodes in the network and stores this information into the routing table. Then it sends this table to all its neighboring nodes. When a neighboring node receives this table, it updates its own table to reflect any changes and broadcast the changed part to others. Therefore, every node can know the delay between it and all the other nodes in the network by DBF. Although DBF is popular in wired networks, it cannot be directly used by wireless ad hoc networks due the routing loop problem [2]. So it cannot be used directly in WMNs either. Here the dynamic source route protocol (DSR) [5], [16], which is popular in wireless ad hoc networks, is adopted for WMNs. The major effort of our proposal is how to integrate the above

2234

TABLE I TABLE AT NODE 2 FOR DELAYS BETWEEN TWO NODES source 1 3 2 2 3 4 ......

destination 2 2 1 3 4 3 .......

delay (s) 0.0055 0.0036 0.0052 0.0045 0.0086 0.0052 .......

0

1

2

3

4

5

6

7

8

9

a) Chain topology 14

7 20

25

mentioned DBF algorithm into DSR in order to obtain the information to estimate the delay over the remaining journey of a data packet for the DQS implementation. This is also because DSR does not broadcast routing packets periodically but routing control packets when there are errors in a link. Therefore it cannot exactly reflect the recent delay to be experienced by data packets in the network through using control packets to probe the delay. Now we discuss how to combine DBF and DSR so that the combination can periodically generate a hello probe packet to estimate the delay between two immediately neighboring nodes. Assume that there are five nodes in line from node 1 to node 5. Node 1 sends out a hello packet at time α, which is received by node 2 at time β. Then the delay from node 1 to node 2 is β − α. Now node 2 can create a delay table, which is used to store delays experienced by packets sent by its immediate neighbors to it, and the same for the other nodes. In order to let node 2 know the delay between other nodes such as between nodes 3 and 4 or between nodes 4 and 5, these nodes must broadcast their delay tables when they send hello packets. After a period of time, node 2 can get the delay between any two nodes as shown in Table I. Given this table, the scheduler looks up the table to estimate the delay over the remaining journey of a packet upon its arrival and then queues it accordingly. 2) Queueing control packets: This issue can be relatively easily resolved by using a dedicated queue to store control packets as illustrated in Fig. 1. This queue will be always given priority over the data packet queue and the packets in this queue will be serviced in the fashion of first-in-firstout (FIFO). The control packets here include route broadcast packets, address resolution protocol (ARP) packets, ACK control packets for TCP, etc. III. S IMULATION STUDIES In this section, we use NS-2.29 [8] as the the simulation platform. The simulation model developed by the CMU Monarch project for for ad hoc networks is also adopted for wireless mesh networks here. This model includes a physical layer, an 802.11 MAC layer and a data link layer. In each simulation, a constant bit rate (CBR) source and a variable bit rate (VBR) source are used as traffic sources. The default size of the buffer for each queue is 50 packets. Two network topologies, chain and random as illustrated in Fig. 2, are simulated by increasing the number of hops and sources. Due

12

0

16 13

8 21 23

11

4 19

1 10

2 24

6

15

17

3

5

22 18

9

b) Random topology

Fig. 2.

Network topologies simulated with NS2: chain and random

to similarities between EDF and DQS mentioned previously, we compare our proposal with EDF. We use the following performance metrics to evaluate the two algorithms. 1) Average goodput: the ratio of the number of data packets received by the destination node to the number of data packets transmitted by the source node. Here, we only count those packets whose end-to-end delays satisfy the pre-defined value. 2) Average end-to-end delay: the average delay for a packet to travel from a source node to its destination node, which includes route discovery time, per-node queueing delay, MAC-layer delay and transmission time in the wireless channel as well as propagation delays [1]. We first look at the chain topology as illustrated in Fig. 2(a) by evaluating the scheduling algorithms against the number of hops. There are two traffic flows which are generated by the first node. The simulation time is 500s. The end-to-end delay of CBR is set to 1s and the end-to-end delay of VBR is set to 1.5s. Fig. 3 shows the average goodput of the DQS and EDF scheduling algorithms. For CBR traffic as illustrated in Fig. 3(a), when the number of hops is less than or equal to 6, the destination node can receive all the packets sent by the source within the predefined end-to-end delays for both DQS and EDF. Once the number of hops is larger than 6, both algorithms decrease their goodput because the delay for some data packets cannot be guaranteed as the number of hops increases. As mentioned earlier, DQS can consider the endto-end path situation of applications. It can schedule packets according to the number of hops in a path, but EDF just

2235

CBR−UDP 1

DQS EDF

1

VBR−TCP DQS EDF

DQS EDF

1

DQS EDF

0.9

0.95

0.9

average goodput

average goodput

0.95

CBR−UDP

VBR−TCP 1

0.9 0.85 0.85 0.8 0.8

0.8

0.95

0.7 0.6

0.9

0.75

0.65

0.5

0.75

0.7

2

4

6 8 hops

Fig. 3.

10

12

0.7

2

4

0.4

12

DQS EDF average e2e delay (s)

average e2e delay (s)

0.3

0.2 0.2

VBR−TCP

2

Fig. 4.

4

6 8 hops

10

12

0

0.14 0.12 0.2 0.1 0.15

0.08

0.1 0.05

2

4

6 8 hops

10

12

0.06 5 10 15 total sources in network

Fig. 6.

Average end-to-end delay for the chain topology

schedules packets based on a given per-hop delay and does not consider the end-to-end path situation. So EDF is more sensitive than DQS when the number of hops increases. For VBR traffic as illustrated in Fig. 3(b), the phenomenon is almost similar to that for CBR traffic except that EDF is slightly better that DQS when the number of hops is less than or equal to 6. As illustrated in Fig. 4, the average delay of DQS is less than that of EDF due to the same reason as mentioned earlier. That is, DQS takes the end-to-end path as scheduling parameter so that it can get less delay than EDF. However, there is a V-shape on average delay when the number of hops is 6. This phenomenon is related to the simulated topology since the similar phenomena also exists in other scheduling algorithms such as CMUPriQueue and Droptail provided by the simulation package. We carried some simulation due to the space limitation, the results are not presented here) by using these algorithms and found that there is a MAC multiplexing problem, which is related to hidden and exposed-terminals.

DQS EDF

0.18

0.25

0.05 0

0.2

0.16

0.3

0.1 0.1

5 10 15 total sources in network

Average goodput for the random topology

DQS EDF

0.35

0.3

0.15

0.85

CBR−UDP 0.4

0.4

0.25

5 10 15 total sources in network

Fig. 5.

VBR−TCP 0.5

DQS EDF

0.35

10

Average goodput for the chain topology

CBR−UDP 0.4

6 8 hops

0.04

5 10 15 total sources in network

Average end-to-end delay for the random topology

The middle node just divide this chain topology into two noninterfering parts and each has two hops. For the random topology, we used 20 terminals and 6 mesh routers in a square grid with a dimension of 800m × 800m as illustrated in Fig. 2(b). We vary the traffic load by changing the number of sources from 2, 4, 8, 12 to 16. The numbers of CBR and VBR sources are the same for every simulation. Fig. 5 shows the average goodput for these two scheduling algorithms. Due to the effect of other nodes in network, for example, hidden and exposed terminals, the average goodput here is not as high as that in the chain topology. The decrease of the average goodput for EDF is more obvious than that for DQS as the number of sources increases. As discussed for Fig. 3, DQS can queue packets according to end-to-end path situation so that it can get higher average goodput than EDF. As illustrated in Fig. 6, changing the number of sources has little influence on the average end-to-end delay. This is because end-to-end delay has a closely relationship with the length of end-to-end path. In our simulation, the length of endto-end path for applications is roughly smaller than 6 since it

2236

is difficult to control the path length with a random topology. This results in that the difference on average end-to-end delay is not so obvious. IV. C ONCLUSIONS AND FUTURE WORK In this paper, we investigated how to use the differentiated queueing service (DQS) scheduling algorithm that was originally proposed for wired networks in wireless mesh networks (WMNs). We proposed a cross-layer implementation of DQS for WMNs, and compared it with EDF through simulation studies. In general, the results show that DQS can outperform EDS in the environments we simulated. There are several interesting issues to be addressed in the future such as how to more efficiently handle these packets whose end-to-end delays are perceived as “not able to be guaranteed”. There are many options available to handle this issue, such as simply dropping or queueing them into a besteffort queue etc, which require further studies to find better solutions. Another important issue is how to consider terminal mobility in the estimation of delay over the remaining journey to the destination of packets for DQS implementation.

[13] S. Lu, V. Bharghavan, and R. Srikant, “Fair scheduling in wireless packet network,” IEEE/ACM Trans. Networking, Vol. 7, No. 4, pp. 473-489, AUGUST 1999. [14] S. M. Jiang, “Granular Differentiated Queueing Services for QoS: Structure and Cost Model,” ACM SIGCOMM Computer Communication Review, Vol. 35, No. 2, pp. 13-22, April 2005. [15] Y. Cao and V. O. K. Li, “Scheduling Algorithms in Broad-Band Wireless Networks,” IEEE proceedings of the IEEE, Vol. 89, No. 1, pp 76-86, January 2001. [16] Y. Mohanmad and Y. Qian, “Performance Study of a Secure Routing Protocol in Wireless Mobile Ad Hoc Networks,” In Proc. of the 2nd International Symposium on Wireless Pervasive Computing, San Juan, USA, pp. 425-430, Februray 2007.

ACKNOWLEDGMENT We would like to take this opportunity to sincerely thank the anonymous receivers for their constructive comments and suggestions. Also thank very much Ms Karen Mohn of IEEE for her support and patience for helping us to correct our initial wrong conference registration. R EFERENCES [1] B. G. Chun and M. Baker, “Evaluation of Packet Scheduling Algorithms in Mobile Ad Hoc Networks,” ACM Mobile Computing and Communications Review, Vol. 6, No. 3, pp. 36-45, July 2002. [2] C. E. Perkins and P. Bhagwat, “Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for Mobile Computers,” In Proc. of ACM SIGCOMM, London England, UK, pp. 234-244, August 1994. [3] C. Fragouli, V. Sivaraman, and M. Srivastava, “Controlled multimedia wrireless link sharing via enhanced class-based queuing with channelstate dependent packet scheduling,” Proceedings of INFOCOM, Vol. 2, pp. 572-580, March 1998. [4] D. Bertsekas and R. Gallager, “Data Networks,” Englewood Cliffs, NJ: Prentice-Hall, 1987. [5] D. Johnson, D. A. Maltz, and Y. C. Hu, “The dynamic source routing protocol for mobile ad hoc networks,” IEFT Internet draft, Mobile Adhoc Network Working Group, IEFT, March 2001. [6] H. Zhang, “Service Disciplines for Guaranteed performance Service in Packet-Switching Networks,” IEEE proceedings of the IEEE, Vol. 83, No. 10, pp. 1374-1396, October 1995. [7] I. F. Akyildiz and X. Wang, “Wireless mesh networks: a survey,” Computer Networks, vol. 47, no. 4, pp. 445-487, september 2005. [8] K. Fall, K. varadhan, and editors, “ns notes and documentation,” The VINT Project, UC Berkely, LBL, USC/ISI, and Xerox PARC, July 1999. [9] L. Georgiadis, R. Guerin, and A. Parekh, “Optimal multiplexing on a single link: Delay and buffer requirements,” IEEE Transactions on Information Theory, Vol. 43, No. 5, pp. 1518-1535, September 1997. [10] N. Nandiraju, D. Nandiraju, L. Santhanam, B. He, J. F. Wang, and D. P. Agrawal, “Wireless Mesh Network: Current Challenges and Future Directions of Web-in-the-sky,” IEEE Wireless Communications Magazine, Vol. 14, Issue 4, pp. 79-89, August 2007. [11] R. Bruno, M. Conti, and E. Gregori, “Mesh Networks: Commodity Multihop Ad Hoc Network,” IEEE Communications Magazine, vol. 43, no. 3, pp. 123-131, March 2005. [12] R. Guerin and V. Peris, “Quality-of-service in packet networks: Basic mechanisms and directions,” Computer Networks, vol. 31, no. 3, pp. 169-189, February 1999.

2237

Suggest Documents