Quality of Service in Wireless Networks Based on Differentiated Services Architecture

Quality of Service in Wireless Networks Based on Differentiated Services Architecture Indu Mahadevan and Krishna M. Sivalingam 1 School of Electrical Engineering and Computer Science, Washington State University, Pullman, WA 99164 Abstract This paper considers the problem of providing signaling support for quality of service (QoS) in wireless and mobile networks. In particular, it proposes an architecture based on Differentiated Services (Diffserv) framework being developed for wired networks. The architecture presents the need for an explicit signaling protocol that is necessary since Diffserv uses an implicit mechanism. The Class Based Queuing (CBQ) algorithm has been used for scheduling user requests and has also been modified to handle several wireless specific features. This includes user mobility, packet loss characterization, lower wireless bandwidth, and battery power constraints. The framework and mechanisms have been implemented in a wireless testbed for the FreeBSD operating system using Pentium workstations and WaveLAN wireless equipment. Experimental results from this testbed show the validity of the proposed Diffserv model and also provide performance analyses. I. INTRODUCTION There is a growing need to provide Quality of Service (QoS) for mobile and wireless applications due to increase in real-time applications and availability of wireless networks. Provision of QoS has been studied at various levels in the protocol hierarchy. These include: (i) MAC-level reservation protocols such as EC-MAC [1], (ii) Scheduling algorithms for wireless networks that consider channel state, losses, etc. [2], and (iii) Network-level signaling protocols such as Resource Reservation Protocol (RSVP) [3]. This paper considers an alternate mechanism to the RSVP signaling protocol since the latter is known to have scalability constraints. The QoS architecture proposed in this paper is based on the Differentiated Services (Diffserv) [4] model. Diffserv uses the Type of Service (TOS) field of the IP packet header for classification. Each IP packet is then given a particular forwarding treatment at each node resulting in some form of quality assurance. Diffserv has better scalability properties compared to the Integrated Services (Intserv) model that uses flow-level QoS specifications. However, Diffserv does not provide the fine-grain guarantees that is possible with Intserv. Although Diffserv is in the relatively early stages of design and definition, simulation studies and some implementations have shown promising results. The current design of Diffserv considers only wired networks, and we propose an enhanced Diffserv framework with modifications to make it suitable for wireless networks. In particular, enhancements have been added to address factors 1 Corresponding Author. Part of the research was supported by Air Force Office of Scientific Research grants F-4962-97-1-471 and F- 4962-99-1-125. E-mail: fimahadev,krishnag@eecs.wsu.edu. including signaling, user mobility, battery power constraints, high link losses, and low bandwidth. The proposed framework is then implemented and tested on a wireless testbed. The testbed consists of Pentium based workstations and Lucent WaveLAN wireless network equipment, and is developed for the FreeBSD operating system. The experimental results show: (i) the need for a signaling/messaging protocol as opposed to the implicit admission control of Diffserv networks, (ii) the need for a performance feedback mechanism due to the high losses in wireless network, (iii) the need for periodic power profile information and the characterization of a packet scheduling behavior that uses this power profile, and (iv) the need for signaling and bandwidth allocation when the user moves between regions. The goal of this work is to develop a generic Diffserv framework for wireless, upon which other specific mechanisms may be implemented. The paper is organized as follows. Section II describes the architectural framework and the factors affecting design. The testbed and implementation details follow in Section III. Experimental results and conclusions follow in Sections IV and V respectively. II. QOS ARCHITECTURE FRAMEWORK A campus-based network architecture is considered in the paper. The campus network is assumed to be composed of various departmental subnets that are connected to a campus subnet, which in turn is connected to the Internet (fig. 1). The architecture assumes various regions ( cells as in a cellular architecture) and each region has a base station (BS) serving all mobiles within its region and is connected to the wired network. When a mobile moves to another region, it is handed off to the base station serving that cell. Departmental Network Campus Network R3 Subnet A Subnet B Subnet C BS BS BS Internet Figure 1: Architecture of a Campus Network. R2 R1 R4 BS BS BS

A. Overview of Differentiated Services The Diffserv architecture [4 6] uses the Type of Service (TOS) field in the IP header to classify flows. The architecture is scalable because it does not maintain per-flow state and there is no requirement for end-to-end signaling. In the Diffserv model, the IP TOS byte (or DS byte) is divided into a 6-bit differentiated services code point (DSCP) and 2 unused bits. Services can be constructed by a combination of: (i) setting bits in an IP header field (DSCP) at network boundaries, (ii) using those bits to determine how packets are forwarded by the nodes inside the network, and (iii) conditioning the marked packets at network boundaries in accordance with the requirements or rules of each service. While Diffserv provides a scalable architecture, it is not well suited in its present form for wireless networks because: (i) characteristics of wireless networks are not taken into account, (ii) Diffserv uses implicit admission control, i.e. a sender can send data at any rate and the capabilities of the receiver, which is important in a wireless network, are not considered, and (iii) there is a need for some kind of minimal signaling to take into account mobility and to cater to local conditions in wireless networks such as mobile battery power level. The goal of this paper is to provide an enhanced Diffserv architecture for wireless networks and in particular: (i) Provide a simple, scalable QoS mechanism which is easy to implement, (ii) Provide assurance over a long-term which is advantageous to many applications, (iii) Use a simple signaling mechanism to transfer control information, and (iv) Incorporate the significant features of a wireless/mobile network into the mechanism. B. Design Details In this section, details of the various components used in our architecture are provided, where each component provides a specific functionality. Most of the discussion is focused on the wireless link between the base station and the receiver. Traffic Classifier and Conditioner: For QoS assurances it is not only necessary to allocate a specific bandwidth to a class of application but is advantageous to share bandwidth among various applications. The Class Based Queuing (CBQ) [7] mechanism is used in our architecture as a traffic classifier, scheduler and conditioner. CBQ was designed to provide link-sharing in wired networks among applications. Each node (host or routers) that is capable of QoS control needs a packet scheduler and a classifier, which is handled by CBQ mechanisms. It consists of: (i) a Classifier, which classifies packets into a pre-defined class, (ii) an Estimator, which estimates bandwidth usage of each class and a (iii) Packet scheduler, which selects the packet or class scheduling. Traditionally CBQ classifies packets based on the source and destination addresses and ports. Our architecture will use a modified CBQ implementation which uses the DS byte to classify data. Signaling Mechanism: Diffserv architecture uses an implicit admission control mechanism. In wireless networks a simple signaling scheme would be required and advantageous because: (a) static provisioning alone is not enough when mobility is concerned, the sender must know the limitations of the wireless link for a better performance and (c) various information on local conditions like power status of the mobile etc. need to be sent occasionally between the base station and the mobile. The signaling protocol proposed: (i) is simple and scalable, (ii) does not maintain any state, and (iii) does not require introduction of any new protocol. We have modified the Internet Control Message Protocol (ICMP) [8] for signaling. Various new ICMP types and codes are used to denote various signaling messages that are sent between the mobile and the base station. Mobility: With mobile networks, it is implied that mobiles often enter and leave a particular cell (region). This means that the base station must accommodate the new mobile s bandwidth requirements. When a mobile moves into the region of a base station there is a possibility that it might be refused bandwidth. A class called new-mobile with a certain amount of link bandwidth is created and new mobiles are allocated bandwidth from this class. When no mobile is using this bandwidth, it is distributed among the various classes that need extra bandwidth. Any new mobile that enters a base station will use the new-mobile class for a specified amount of time. After this period the base station finds out about the nature of the application and modifies the priority of the class based on the availability of bandwidth. Priority is given to systems that are already in the cell. An alternative approach is re-allocate bandwidth from low priority applications of mobiles that are already in the cell to the new users. Both approaches have been implemented and studied. High Loss Rate Handling: Wireless networks are characterized by more frequent packet losses. This necessitates: (i) to be able to specify a nature of packet loss, and (ii) to provide some kind of a feedback and compensation for the data being received. The nature of packet loss can be specified with the use of the loss-profiles parameter [9]. The feedback/compensation mechanism can be incorporated by occasionally receiving information about the data received using the rate feedback factor. The feedback from the receiver calls for some kind of compensation from the sender. We propose using a fraction of bandwidth for the compensation class which has its own DSCP. Whenever a feedback is obtained from the mobile, a particular DSCP is compensated by: (i) modifying the policer for the class to include any available compensation bandwidth or (ii) by re-marking some packets of the flow to the DSCP of the compensation class. Both these approaches have their advantages and disadvantages discussed in detail in [1]. The policies for compensation can use the approaches discussed in [11, 12].

Low Bandwidth Handling: Since Diffserv uses implicit admission control the receiver does not have a way to indicate whether it cannot support this bandwidth. A signaling mechanism must be used to overcome this. Once the capabilities of the wireless link are known using signaling there is a need to restrict the bandwidth. The bandwidth restriction can be done at two levels: (i) at the sender (a wired host) or/and (ii) at the base station. In our testbed, we use the second approach in case of a point-to-multipoint applications, and the first approach is used for other applications. Power Constraints Handling: A mobile is characterized by power restrictions because of the use of limited power batteries. Some of the major consumers of power in a mobile are the network interface (14%) and the CPU/memory (21%). A mechanism is proposed that uses the current battery power level in scheduling. The power level is periodically sent to the base station using the signaling mechanism. The scheduling of packets is done with the use of the power profile parameter. The power profile parameter identifies the nature of the application. For example, for layered video, the power profile can dictate that only the base level packets be sent. For other applications, the average rate of the data sent can be reduced based on the power profile. III. TESTBED IMPLEMENTATION In the previous section the various enhancements for wireless networks were discussed. The details of the testbed and some implementation details are provided in this section. A. Testbed Details The testbed has three Pentium systems that operate as base stations. Each base station is equipped with an Ethernet card and a 2.4 GHz Lucent WaveLAN ISA card. The base stations are in adjacent cells and the different Network Identifiers (NwID) of the WaveLAN cards identify the cells. The testbed also has two mobiles which are equipped with 2.4 GHz PCMCIA WaveLAN cards. All the systems run the FreeBSD 2.2.2 operating system. The testbed also uses: FreeBSD WaveLAN driver for PCMCIA cards which supports roaming [13]. The WaveLAN ISA driver has been modified to produce link level beacons to identify a particular cell. The mobile uses the NwID of the beacon signal to identify the base station it is attached to. The signal strength of the beacon helps determine when a mobile is moving from one cell to another. The driver from [13] has also been modified to (i) work with 2.4GHz cards, (ii) use CBQ and (iii) work with our modified WaveLAN ISA driver beaconing system. Modified CBQ that (i) uses DSCP to classify flows, (ii) schedules packets based on power levels and (iii) uses compensation/feedback information and modifies flows accordingly and (iv) uses information from a new mobile to allocate bandwidth. Signaling mechanism based on ICMP to send messages between the sender, base station and the mobile. traffic generator program. Advanced Power Management (apm) [14] tool for monitoring power on laptops. Sender - Wired host Base station Receiver - Mobile report: pwr profile modify: bandwidth modify: bandwidth report: pwr level (a) Figure 2: Flow of Signal Messages in the System. B. Implementation Details Base station report: bandwidth reply: bandwidth report: pwr level New Mobile In this section, the implementation details of a few components in the system are discussed. Implementing the Signaling Mechanism: Three kinds of ICMP signaling messages are defined: (i) Report: which conveys any local parameter like bandwidth requirement, power level, etc. (ii) Reply: to acknowledge any bandwidth reservation request and (iii Modify: conveying the bandwidth changes. The flow of necessary signaling messages along a timeline are shown in fig. 2(a) and. When a session begins, the sender sends the power profile of the application to the base station as shown in fig. 2(a) after which data is sent (not shown in the fig. 2(a)). All solid lines denote the mandatory messages while the dashed lines represent optional messages sent based on the circumstances. Modifications to CBQ: The modifications to CBQ include: (i) classification of packets based on the TOS field in IP header, (ii) packet dropping extensions based on the nature of power profile. Currently the possible power profiles include sending packets based on the differential importance bit in DSCP field and reducing the rate of sending packets, (iii) mechanisms to alter the bandwidth allocation of a class and (iv) mechanisms to map a certain percentage of packets belonging to one class to use another class. Further details of the implementation may be found in [1]. IV. EXPERIMENTAL RESULTS In the previous section, the experimental details of the testbed were discussed. Various features of the architecture and how they perform are presented in this section. A. Reservations Under User Mobility The experiments in this section show two methods of accommodating a new mobile in a cell. In the first case, a certain percentage of the bandwidth is reserved for the new mobiles entering the area. In the second case, a low priority application relinquishes some of its bandwidth for the new application. The experimental setup consists of a sender, two base stations and two mobiles. The working of both the scenarios is described using fig. 3(a) and.

(1) 1.5 (1) User 1.4 1.5 (2) (3) (4) (5) User 1 & User 2 User 3.2 (2) User 3 User 2 new-mobile 1 2 3 4 (a) 1 2 3 Figure 3: Bandwidth Allocation During Mobility: (a) Using bandwidth from the new-mobile class and Using bandwidth of a low priority application. Using Bandwidth of New-mobile Class: In this experiment, User 1, User 2 and new-mobile class are alloted 3%, 3% and 2% of the bandwidth respectively. The default class takes up the rest of the 2% of the bandwidth. When a mobile enters the cell, 15% of the bandwidth from the new-mobile class is alloted to user 3. Fig. 3(a) aids in the discussion here. User 1 and User 2 are initially sending data and since the new-mobile class is not being used, it is alloted to users 1 and 2 (shown as time instant (1) in fig. 3(a)). At instant represented as (2), a mobile enters the region and utilizes the new-mobile class bandwidth and the bandwidth available to users 1 and 2 is reduced. The user from the mobile (User 3) uses the new-mobile class for a while (shown as time instant (3)) after which User 3 is alloted its own class and the bandwidth of the new-mobile class is reduced (shown as time instant (4)). Any extra bandwidth in new-mobile class is borrowed by Users 1, 2 and 3 (shown as time instant (5)). Reducing Bandwidth of a Low Priority Class: In this experiment User 1 and User 2 are alloted 4% of the bandwidth each. As in the previous case 2% of the bandwidth is reserved for the default class. Fig. 3 aids in the discussion here. Initially User 1 and User 2 are sending data at an equal rate (shown as (1) in fig. 3(a)). At the instant marked (2), a mobile has entered the cell with a high priority application. The allotment of User 2 (a low priority application) is reduced to 1% while the new mobile uses up 3% of the remaining bandwidth. B. Compensating Losses Wireless networks are categorized by high loss rates and there may be a need to compensate certain classes with extra bandwidth when necessary. Compensation can be provided in two ways: (i) by providing required bandwidth from compensation class or (ii) by allowing some packets from the flow that needs to be compensated to use the DSCP of the compensation class. The experimental setup consisted of a sender, a base station and a mobile. Both the methods mentioned above are discussed with the use of fig. 4(a) and which denote the cases (i) and (ii) mentioned above respectively. In Fig. 4(a), initially the user class is sending data at a rate of around.23 Mbps when a feedback from the receiver arrives (shown as (1)). The bandwidth of the compensation class is reduced to around.18 Mbps while the user gets this extra bandwidth to send data at the rate of.4 Mbps. After the particular class has been compensated for a while, the user class reverts to its original rate (denoted as (2) in fig. 4(a)). The method of re-marking some packets in a flow to use the compensation bandwidth is shown with the help of fig. 4. When the bandwidth report arrives (shown as instance (1)), a fraction of the packets are remarked to use the DSCP of the compensation class. Effectively between the instant (1) and (2) in fig. 4, the user class is able to send.3 Mbps of data (.2 Mbps from the original allotment and.1 Mbps from the compensation class). C. Receiver Feedback Under Low Bandwidth The experimental setup for feedback from the receiver consisted of a sender (a wired host), a base station and a mobile (receiver). The first experiment shows how implicit admission control affects a low bandwidth wireless environment. The sender sends data at a certain rate without considering the capabilities of the wireless link. Reservation of 2% on the Ethernet side resulted in a class of capacity 1.94 Mbps while a 2% reservation for a class on the wireless side resulted with.24 Mbps. Sender sends the flow at the rate of.425 Mbps. In the experiment, all wired hosts are able to sustain this bandwidth. Fig. 5(a) shows the effect of implicit admission control. In fig. 5(a), a default buffer size of 3 packets is used and the class is not allowed to borrow bandwidth from other classes. We notice that as the queue size increases, the delay for each packet is increased. Once the queue size reaches a maximum limit of 3 packets, packets start getting dropped. At this point we have two alternatives: (i) do some improvements at the base station to accommodate the flow or (ii) inform the sender to cut

User class Compensation Class.4.2 1 2.2.1 User Class Compensation Class 1 2 2 4 6 8 (a) 2 4 6 8 Figure 4: Compensation for Classes with High Losses: (a) Using bandwidth from compensation class and Re-marking packets to use DSCP of compensation class. down on the sending rate. Note that such a scheme is needed for UDP. TCP has its own congestion control mechanism. The improvements at the base station include: (i) increasing buffer size (fig. 5) and (ii) allowing the class to borrow bandwidth from other classes (fig. 5(c)) if available. In the first case, a increase in queue size to 2 2 shows that packets are dropped much later on. This could be enough to sustain some flows. In the second case, borrowing of bandwidth from other classes is turned on. Since the other class was not using its alloted bandwidth, the flow could borrow a major portion of the bandwidth and there is no delaying and dropping of packets (as shown in fig. 5(c)). This is not always possible since the other applications may be using their required bandwidth. In case both these enhancements at the base station are not enough and packets are dropped indiscriminately, a message needs to be sent to the sender to reduce the rate of sending data. Fig. 5(d) shows the data sent by the sender before and after the signal is received. We notice that the number of packets sent are reduced after the signal is received at around 3 time units. D. Power Constraint in Wireless Networks Mobile battery power level is an important constraint in wireless networks and it is therefore advantageous to periodically send the mobile power level to the base station, and let the base station and the mobile adapt the application behavior based on the power profile. In our testbed, the sender initially sends the power profile message to the receiver based on the nature of application it is going to send. This power profile information is used by the base station or the mobile while scheduling in a low power situation. An adaptive video application that changes the type and amount of video information sent was used to test the scheme. This experiment is based on work reported in [15]. Consider an MPEG-1 and MPEG-2 video stream with I (Intra), B (Bidirectional) and P (Predictive) frames. The idea is to drop B frames first because this affects only the particular frame and the decoder can extrapolate the rest of the message. If the power level is not achieved then P frames are discarded. I frames are discarded last. We used a simulated video stream 2 The maximum queue size that could be tested was 2 due to practical memory limitations. which uses the ratio of I, P and B frames for two video rates of 1 Mbps and 384 Kbps as obtained from [16]. The power consumed at the mobile was measured using apm (advanced power management) [14] tool. The experiment did not produce a very definitive result because we do not currently have a tool to accurately measure the battery power consumption on a per-packet basis. The apm tool just reports the percentage of battery power left which is a coarse measurement. Various other power measuring mechanisms are being considered. In this section, the experimental results from the various enhancements were shown. Section V concludes the paper. V. CONCLUSIONS This paper addressed the various enhancements that need to be made to Diffserv to make it suitable for providing QoS assurances in wireless networks. Various characteristics of the wireless networks like high loss, low bandwidth, battery power constraints and mobility were considered in tailoring Diffserv. Experimental results used to validate the enhancements include: (i) testing of the CBQ-based scheduling scheme for Diffserv, (ii) mechanism to provide compensation bandwidth in case of high loss in wireless networks, (iii) use of signaling mechanism for improving performance in low bandwidth situations, (iv) power level information for power-intelligent scheduling, and (v) bandwidth allotment alternatives during mobility. We point out that this paper was concerned with a general overall enhancement of Diffserv and did not concern with any particular per-hop behavior (PHB). A look into various possible PHBs for wireless networks will be very useful. VI. REFERENCES [1] J.-C. Chen, K. M. Sivalingam, and R. Acharya, Comparative analysis of wireless ATM channel access protocols supporting multimedia traffic, ACM/Baltzer Mobile Networks and Applications, pp. 293 36, 1998. [2] C. Fragouli and V. Sivaraman and M. Srivastava, Controlled multimedia wireless link sharing via enhanced class-based queuing with channel-state-dependent packet scheduling, in Proc. IEEE INFOCOM, (San Francisco, CA), pp. 572 58, Apr. 1998.

(a) 4 3 35 Delayed packets Packets dropped Queue Size 25 Delayed packets Packets dropped Queue Size 3 2 25 2 15 15 1 1 5 5 1 2 3 4 5 6 25 Total borrowed Total packets 1 2 3 4 5 6 7 18 16 Total packets 2 14 12 15 1 1 8 6 5 4 2 1 2 3 4 5 6 7 1 2 3 4 5 6 7 (c) Figure 5: Performance of Low Bandwidth Wireless Link: (a) Effect of implicit admission control, Effect of increasing queue size, (c) Effect of borrowing bandwidth, and (d) Informing sender of wireless link limitations. (d) [3] I. Mahadevan and K. M. Sivalingam, Architecture and Experimental Results for Quality of Service in Mobile Networks using RSVP and CBQ, ACM/Baltzer Wireless Networks, July 1999. (Accepted for Publication). [4] Y. Bernet, J. Binder, S. Blake, M. Carlson, E. Davies, B. Ohlman, D. Verma, Z. Wang, and W. Weiss, A Framework for Differentiated Services. IETF draft, draft-ietf-diffservframework-.txt, May 1998. [5] K. Nichols, S. Blake, F. Baker, and D. L. Black, Definition of the Differentiated Services Field in the IPv4 and IPv6 headers. IETF draft, draft-ietf-diffserv-header-4.txt, Oct. 1998. [6] V. Jacobson, K. Nichols, and K. Poduri, An Expedited Forwarding PHB. IETF draft: draft-ietf-diffserv-phb-ef-.txt, Aug. 1998. [7] S. Floyd and V. Jacobson, Link-sharing and resource management models for packet networks, IEEE/ACM Transactions on Networking, vol. 3, pp. 365 386, Aug. 1995. [8] J. B. Postel, Internet Control Message Protocol. rfc792, Sept. 1981. [9] K. Seal and S. Singh, Loss Profiles: A quality of service measure in mobile computing, ACM/Baltzer Wireless Networks, vol. 2, pp. 45 61, 1996. [1] I. Mahadevan and K. Sivalingam, Differentiated Services for Quality of Service in Wireless networks: Architecture and Implementation, Tech. Rep. TR99-DAWN-3, School of EECS, Washington State University, Email: krishna@eecs.wsu.edu, Aug. 1999. [11] P. Ramanathan and P. Agrawal, Adapting Packet Fair Queuing Algorithms to Wireless Networks, in Proc. ACM MobiCom, (Dallas, TX), pp. 1 9, Oct. 1998. [12] S. Lu, T. Nandagopal, and V. Bhargavan, A Wireless Fair Service Algorithm for Packet Cellular Networks, in Proc. ACM MobiCom, (Dallas, TX), pp. 1 2, Oct. 1998. [13] Wavelan driver on FreeBSD supporting roaming. http://www.monarch.cs.cmu.edu/wavelan.html. [14] Advanced Power Management. http://sunsite.unc.edu/ LDP/HOWTO/mini/Battery-Powered-3.html. [15] P. Agrawal, J.-C. Chen, S. Kishore, P. Ramanathan, and K. M. Sivalingam, Battery Power Sensitive Video Processing in Wireless Networks, in Proc. IEEE PIMRC 98, (Boston, MA), Sept. 1998. [16] J. M. Boyce and R. D. Gaglianello, Packet Loss Effects on MPEG Video Sent Over the Public Internet. www.acm.org/sigmm/mm98/electronic proceeding/boyce /index.html#intro, 1998.