QoS Featured Wireless Virtualization based on Hardware

Transcription

1 QoS Featured Wireless Virtualization based on Hardware Cong Wang and Michael Zink Department of Electrical and Computer Engineering University of Massachusetts, Amherst, MA {cwang, Abstract-Nowadays wireless network virtualization technique provides a practical perspective for splitting a single wireless network into multiple independent slices to support isolated virtual networks. A key problem for infrastructure providers is the fairness that needs to be provided among different virtual networks. In this paper, we provide a simplified method implemented by Click Modular router to take advantage of wireless virtualization and provide QoS featured airtime fairness for various services based on the application priorities of the users. Practical experiments show that our bandwidth shaping mechanism can provide more flexible resource allocation based on user needs and stable throughput guarantee for QoS featured services. Index Terms-Network measurement, Quality of Service, Wireless network virtuailization, Streaming media I. INTRODUCTION In recent years the virtualization of computer networks has gained tremendous popularity. It allows the creation of multiple (virtual) networks, where each one might serve a different purpose, simultaneously over the shared substrate (physical) network. While initial network virtualization mainly focused on wired networks, there has recently been research in expanding the virtualization concept into wireless network fields. In this paper, we show a virtualization method for wireless networks that is based on splitting a single Access Point (AP) into multiple independent ones to create virtual wireless networks (VWLAN) that can serve different end user needs. Our goal is to use virtualization to isolate real-time from non-real-time traffic by allocating appropriate amounts of physical network resources for each virtual network slice. With this approach, real-time data can be given higher priority to ensure a certain level of quality of service (QoS). In previous work [5], we have shown that under current network condition, e.g., campus WLAN, commercial DSL or cable Internet that due to the lack of prioritization mechanisms for real-time data, it is common that users experience playback pauses when streaming videos from major online video websites. Since video and voice service quality largely depends on the available throughput, low throughput can significantly decrease the video or voice quality or even cause service interruption. On the other hand, slightly reducing the throughput of some applications such as file download will not influence the actual content. To solve this problem, we propose the WifiQoS mechanism that is implemented using Click Modular router software architecture [6]. This approach enables wireless network virtualization by splitting a single physical AP into several virtual ones and perform priority based traffic shaping depending on destination and type of the incoming data. Via a simple mechanism of giving Video on Demand services like YouTube and Netflix higher bandwidth, we demonstrate that the quality of service will significantly increase if these wireless clients do not have to compete for bandwidth with other applications that are less throughput sensitive. We evaluate our proposed wireless network virtualization approach in an experimental setup that is part of our university network. To evaluate our approach we make use of two metrics. The first one is PSNR (Peak Signal-to-Noise Ratio), which is a common metric to determine the reduction in video quality between a transmitted video and its original. The second metric is the accumulative pause time which represents initial and intermediate play out delay due to insufficient buffer level. The systematical measurement of relationship between QoS performance and network throughput can be the foundation for further developing an algorithm to dynamically prioritize QoS applications via taking advantage of allocating network bandwidth in real-time. In addition, we compare our work with the existing IEEE 802.lIe standard which is the currently used mechanism for ensuring QoS in networks. Our results show that our method provides higher flexibility, isolation and programmability since more controlling privilege is given to the management node (AP). The remainder of this paper is organized as follows. In Section II, we present related work. We present the WifiQoS design and system implementation in Section III. The measurement results are shown in Section IV. We compare the performance of our work and the existing IEEE 802.lIe standard in Section V. Finally, we summarize our work and point out future works in Section VI. II. RELATED WORK The topic of wireless network virtualization has been investigated intensively in recent years. In [8], the authors focus on virtualization based on Time Division Multiplexing (TOM), and to develop and demonstrate techniques of TOM for slices on the ORBIT facility to determine how different experiments can co-exist in the same wireless facility and explore specific requirements for such co-existence. In [4], a software based /12/$ IEEE 801

2 approach named MultiNet to facilitate simultaneous connections to multiple networks by virtualizing a single wireless card is proposed. The wireless card is virtualized by introducing an intermediate layer below IP, which continuously switches the card across multiple networks. The goal of the switching algorithm is to be transparent to the users who see their machine as being connected to multiple networks. However, none of these research works provide sufficient QoS support for real time network applications. As a means to support the quality of service for wireless LAN applications, the IEEE e standard was approved as an addition of existing IEEE standard family. There has been many investigations focusing on the measurement studies for the 802.lie standard, e.g., [7]. This standard is an important improvement for delay-sensitive and real time applications. One major disadvantage of e is its lack of flexibility and manageability due to its coarse-grained resource allocation mechanism. In these previous works, although different mechanisms have been devised to wireless network virtualization, QoS based resource allocation for virtual networks has not been discussed which is what we explore in this paper. In addition, our mechanism is able to give different priorities to real time applications whose quality decreases in the case of insufficient network bandwidth. III. WIFIQOS DESIGN OVERVIEW AND IMPLEMENTATION In this section, we present the overall WifiQoS design and the implementation. For the AP splitting and air-time bandwidth shaping mechanism, our infrastructure tries to enforce the QoS services to get higher bandwidth and give ISPs the privilege of controlling bandwidth allocation throughout every virtual network. A. Experimental environment Setup Both the AP node and the client nodes are equipped with Atheros Wireless LAN cards, and these nodes run Ubuntu Linux with kernel version , which is required by the Click kernel module. The radio card driver used for our experiments is MadWifi [2], which is an open source Linux kernel device driver for Atheros-based Wireless LAN cards. The Madwifi driver allows the splittig of a single AP into several virtual APs on top of the same physical device to support multiple virtual networks concurrently. Each virtual AP corresponds to a slice of an underlying physical device, which is created when the driver is loaded. Using this property, we setup our measurement topology which is shown in Figure 1. As can be seen, the single AP is split into two virtual APs, each of them is wirelessly connected to a client or a group of clients. In our specific case, the goal is to allocate sufficient resources for the virtual network VWLAN 1 that carries the video traffic to allow video streaming with a high QoS. All the experiments were carried out inside our laboratory with a single physical AP that is connected to the public Internet. The AP node can support as many applications or data streams as the number of configured virtual APs. In our experiments we are supporting two virtual APs and hence two different applications can perform simultaneously. B. Metrics for measurement To measure the performance of our virtualization approach we make use of two commonly used metrics, PSNR and total pause time. The definition of both of the metrics are presented as follows: PSNR: The PSNR stands for Peak signal-to-noise Ratio, which is a common metric for comparison of the quality between an original video and its post transmission version. It represents the ratio between the power of the original data before transmission and the power of corrupting noise that is introduced by compression or loss during transmission. The PSNR is expressed in terms of logarithmic decibel scale. We use this metric to measure the QoS under various network conditions. For our measurements, we use several local media as the reference video, upload the videos to YouTube, compare the streamed media and calculate the PSNR. Total pause time: The main task for service providers is to provide end users smooth video playback which means each byte of the video should arrive at the receiving node before the time it is scheduled to be played. Thus the video player needs to stream a video into the buffer before actual playing the video content, this process is named initial pause time. Most of the real time video players will pause playing when there is insufficient data at the user side buffer, this process is named playback pause time. We calculate the total pause time as the sum of initial pause time and playback pause time. In general, the higher the throughput, the shorter the total pause time. IV. EXPERIMENTAL EVALUATION File Trans"" """"_ Application Fig. 1: General architecture: splitting single wireless access point into multiple virtual access points to supporting different clients In this section, we evaluate the performance of WifiQoS mechanism. We have uploaded a set of videos (30 seconds length each) in different resolutions (varies from 240p to 1080p) to YouTube. All the measurements will be based on this set of media. We begin the measurements with a single client scenario. 802

3 A. Single AP with single client This scenario provides the base case for the evaluation of our WifiQoS approach. We setup a single physical AP, and a single client which wirelessly connects to the AP. We stream videos from YouTube and evaluate the PSNR and total pause time. Since there is no contention from other clients, the client node can get the maximum possible bandwidth and thus get the highest possible video quality iii 24 :s!- o: z 22 (f) PSNR p 360p 480p 720p 1080p Video resolution Fig. 2: The video quality (PSNR) of single client As shown in Figure 4, the PSNR decreases by approximately 1.5 db for each video quality while the throughput of the Iperf-based data transmission increases from 0 to 2 Mbps. It is known that a PSNR change larger than 0.5 db can be noticed by viewers (a variation larger than 2dB will be clearly observed), which is shown in [3]. Thus, the increasing competition from other applications can significantly affect real time video quality. The influence of competing traffic on the pause time can be clearly observed from Figure 5. As the transmission rate for client 2 goes up, the total playback pause time increases very quickly. E.g., the 1080p video total pause time is seconds, which is more than 112 of the overall length of the video (30 seconds). The results show that the real time application users are much more throughput sensitive than general purpose file transmission application users, especially when they watch videos with higher resolution. The long pause time for a video is undesirable, and methods for appropriate resource reservation for the real time VWLAN are needed. 3.5 Pause time I Throughput of client 2 (Mbps) 240p 360p 480p nop 1080p Video resolution Fig. 4: The PSNR of client 1 before bandwidth shaping Fig. 3: The total pause time of single client The total pause time and PSNR for the single client scenario are shown in Figure 2 and Figure 3. In Figure 2, the x axis represents the video resolution, and the y axis represents the video quality in terms of PSNR measured in db. In Figure 3, the x axis represents the video resolution and the y axis represents the pause time (the total of initial and intermediate playback pause time) measured in seconds. As can be observed, the PSNR significantly increases when we change the quality of the video that is streamed from the YouTube server, and while the video quality increases, the pause time is also increasing. This clearly shows the trade-off between high resolution and smooth playback. B. Two virtual APs without bandwidth shaping In this set of experiments, we show the performance when data are streamed from the Internet via two VWLAN to two clients. The scenario for this experiment is shown in Figure 1. Client 1 (connected to VWLAN 1) receives a video stream from YouTube while client 2 (connected to VWLAN 2) receives data transmitted by an remote sender (using iperf tool [1]) that is connected to our lab's wired network. r r---- Throughput of client 2 (Mbps) Fig. 5: The pause time of client 1 before bandwidth shaping C. Virtual wireless network resource allocation through bandwidth shaping In this set of experiments, we show how PSNR and pause time change if our traffic shaping approach is employed to give higher priority to data destined for VWLAN 1. Besides the use of traffic shaping the experiment setup is identical to the one described in Section IV-B. For the non-real-time traffic 803

4 Iperf is configured to transmit at a rate of 8Mbps to ensure resource contention. We investigate the average PSNR when given different link bandwidths and the results are shown in Figure 6. We also record and plot the total pause time as shown in Figure 7. As can be observed, the bandwidth shaping technique provides the needed throughput for video streaming application while slightly reducing the throughput for general file transmission applications. For most video streaming applications, reserving I Mbps link rate can satisfy even 1080p video streaming, with small pause time and higher PSNR. 32,- ---r----,----, _ p G 480p K p ---& p I I I I I I I I L- -L L- -L Researved throughput for client 1 (Mbps) Fig. 8: The PSNR of video streaming application after bandwidth shaping in environment with interference 34,----,----,----,----,----, p 0 480p K p p ,----,----,----,----,----,----, 20 v. 360p G '", 480p x-.... v. nop BOp ---v I I I I I I I I I I I " I I I I 16 L ResealVed throughput for client 1 (Mbps) Fig. 6: The PSNR of video streaming application after bandwidth shaping Reserved throughput for client 1 (Mbps) Fig. 9: The total pause time of video streaming application after bandwidth shaping in environment with interference ,----,-----,----,-----, p G 480p x 720p p... o Reserved throughput for client 1 (Mbps) Fig. 7: The pause time of video streaming application after bandwidth shaping D. Bandwidth shaping with channel interference One important aspect for wireless network virtualization is the isolation from channel interference of other wireless networks. In this section, we measure the performance of WifiQoS in an environment with interference. We set up two additional machines located close to the machines of our experiment setup, of which one machine acts as AP node and the other one acts as client node. Both machines are set onto the same wireless transmission channel as the WifiQoS. We send constant traffic from the AP node to the client node with a transmission rate of 1.5 Mbps, and repeat the experiment of WifiQoS as described in Section IV-C. The bandwidth shaping performance under channel interference is shown in Figure 8 and Figure 9. As can be observed, 1.2 although the PSNR decreases due to the interference of other wireless transmission, the video quality could still increases steadily while the reserved bandwidth increases. Also as shown in Figure 9, it is easy to observe the significant decrease of pause time while increasing the reserved throughput for video streaming, thus, the WifiQoS bandwidth shaping mechanism can dynamically allocate necessary network bandwidth to real time applications and thus steadily improve the QoS while in an environment with interference. V. PERFORMANCE OF IEEE 802.IIE AND WIFIQOS The IEEE standard itself was amended by e which defines QoS enhancements. 802.lle tries to provide guarantees for QoS real time applications, such as Voice over IP or video streaming by provisioning prioritized MAC layer queues. However, e only provides four categories for service differentiation, which does not offer sufficient flexibility for fine-grained resource allocation, as we will show in Section V-B. In this section, we compare the performance of WifiQoS and IEEE 802.lle protocol. A. The IEEE 802. I Ie implementation The MadWifi driver provides WMMlWME support to prioritize certain types of network traffic [7]. The primary purpose of 802.lle is to protect QoS applications from influence of low priority data. The default 802.lle standard provides four different access categories (BE-Best Effort, BK-Background, VI-Video and VO-Voice, priority from low to high). 804

5 B. Performance evaluation To show the effectiveness of our WifiQoS bandwidth shaping mechanism, we compare the performance of IEEE e and WifiQoS. For this set of experiments, we use the same setting of a single AP with two virtual APs supporting two clients as described in Section IV. Thus, for the 802.lie standard, the WMM featured VWLAN 1 is given higher priority (VI-Video) and VWLAN 2 which is running standard application is given lower priority (BK-Background). And for the WifiQoS technique, we configure the topology as described in Section IV-C. We record the throughput of both clients, and the PSNR for the video streaming client. For VWLAN 2, we send UDP traffic using the iperf tool to client 2 with a constant data rate of 1.5 Mbps eAC:VI e AC: BE -0 - WifiQo$ video... x... VVifiQo$ general ---A--- ensure a stable throughput for real time applications. However, since the only choice of WME for video streaming application is AC (VI-Video), and at this priority level, the maximum throughput (2.5 Mbps) is given to video streaming applications while much less throughput (1 Mbps) is given to the AC (BE-Background) application. Our WifiQoS mechanism gives the management node (AP node) more choices for network resources allocation. Thus, while giving sufficient bandwidth (1 Mbps) to ensure the best result of video streaming application, the remaining network resources (1.5 Mbps) can still be allocated to the file transmission application. The PSNR and total pause time comparison are shown in Figure 11 and Table I. As can be observed, when setting the video streaming application throughput as 1 Mbps, the WifiQoS mechanism ensures that the video streaming application performs almost similar to the case when 802.lie is used (with a difference of 0.64 db PSNR and 0.52 second total pause time). However, much more network resources can be flexibly allocated to other general purposed network applications in the WifiQoS case. o TIme (sees) Fig. 10: The throughput comparison of 802. lie standard and WifiQoS Fig. li: WifiQoS PSNR (WifiQoS) -- PSNR ( e) G 9-0. '" iii'. sa 0: 23.5 z w GoG_a Time (sees) The PSNR comparison of 802.li e standard and Throughput Throughput Mean of P- Pause time (VI) (BE) SNR 802. lie Mbps 1.2 Mbps db 2.31 s Mbps Mbps db 2.83 s WifiQoS Mbps Mbps db 2.51 s Mbps Mbps db 2.35 s TABLE I: The quantitative comparison between performance of IEEE 802.lie standard and WifiQoS The throughput comparison is shown in Figure 10. As can be observed, both the 802.lie standard and WifiQoS can VI. CONCLUSION Wireless Network Virtualization techniques provide important features of flexibility for network users and service providers. In this paper, we provide a new technique for splitting single wireless network AP into multiple virtual ones, in order to provide proper network resources for different end users. We show that videoo streaming applications have low performance when competing with other applications when using standard PSNR metric for video quality evaluation. Thus we propose the WifiQoS bandwidth shaping mechanism using Click Modular Router software architecture to provide higher priority support for QoS featured applications. Our results show that the WifiQoS mechanism outperforms the existing IEEE e standard in the perspective of network resources allocation. REFERENCES [1] Iperf traffic generation tool. [2] MadWifi: Multiband Atheros Driver for Wireless Fidelity. madwifi-project.org/. [3] CHAN, A., ZENG, K., MOHAPATRA, P., LEE, S.-J., AND BANERJEE, S. Metrics for evaluating video streaming quality in lossy IEEE wireless networks. In Proceedings of the 29th coriference on Iriformation communications (Piscataway, NJ, USA, 2010), INFOCOM'IO, pp [4] CHANDRA, R. Multinet: Connecting to multiple IEEE networks using a single wireless card. In in IEEE INFOCOM, Hong Kong (2004). [5] KHEMMARAT, S., ZHOU, R., GAO, L., AND ZINK, M. Watching user generated videos with pre fetching. In Proceedings of the second annual ACM conference on Multimedia systems (2011), MMSys '11, ACM, pp [6] KOHLER, E., MORRIS, R., CHEN, B., JANNOTTI, J., AND KAASHOEK, M. F. The click modular router. ACM Trans. Comput. Syst. (August 2000), [7] MANGOLD, S., CHOJ, S., MAY, P., KLEIN, 0., HIERTZ, G., STIBOR, L., POLL CONTENTION, C., AND POLL, F. IEEE 802. lie wireless LAN for quality of service. [8] SMITH, G., CHATURVEDI, A., MISHRA, A., AND BANERJEE, S. Wireless virtualization on commodity hardware. WinTECH '07, ACM, pp