An Improvement of TCP Downstream Between Heterogeneous Terminals in an Infrastructure Network

Transcription

1 An Improvement of TCP Downstream Between Heterogeneous Terminals in an Infrastructure Network Yong-Hyun Kim, Ji-Hong Kim, Youn-Sik Hong, and Ki-Young Lee University of Incheon, 177 Dowha-dong Nam-gu, , Incheon, Korea Abstract. We measured a performance of data transmission between a desktop PC and a PDA in an infrastructure network based on IEEE x wireless LAN. Assuming that a PDA is mainly used for downloading data from its stationary server, i.e., a desktop PC, a PC and a PDA acts as a fast sender and a slow receiver, respectively, due to substantial differences in their computational capabilities. With data transmission between these heterogeneous terminals a transmission time during downstream is slower than that during upstream by 20% at maximum. To mitigate this, we present two distinct approaches. First, by increasing the size of a receive buffer for a PDA the congestion window size of TCP becomes more stable. Thus, an approximate 32% increase in throughput can be obtained by increasing its size from 512 bytes to bytes. Second, a pre-determined delay between packets to be transmitted at the sender should be given. By assigning the inter-packet delay of 5 ms during downstream achieves a best performance which improves by 7% compared to that without such a delay. Besides, such a delay reduces the number of erroneous packets remarkably. Keywords: PDA, TCP downstream, wireless LAN, congestion window, interpacket delay. 1 Introduction The transmission of multimedia over wireless networks using mobile devices is becoming a research topic of growing interest. With the emergence of small wireless handheld devices such as PDAs (Personal Digital Assistants), it is expected that interactive multimedia will be a major source of traffic to these handheld devices [5], [6]. Different types of terminals such as desktop PCs as fixed hosts (FHs) and PDAs as mobile hosts (MHs) can be connected to an infrastructure network. Assuming that a PDA is currently used for downloading data from its stationary server such as a desktop PC, a desktop PC and a PDA acts as a fast sender and a slow receiver, respectively, due to substantial differences in their computational capabilities. Y.-H. Lee et al. (Eds.): ICESS 2007, LNCS 4523, pp , Springer-Verlag Berlin Heidelberg 2007

2 740 Y.-H. Kim et al. Actually, a PDA has a lower performance, less memories and poor user interfaces compared to a desktop PC. Most of the works deal with measurements and analysis of their performance with emphasis on laptop PCs [2], [3]. In that case, a desktop PC and a laptop PC are considered as a fast sender and a fast receiver, respectively, or vice versa. This work provides a performance characterization of three types of transmissions: upstream, downstream and wireless-to-wireless. Measures were carried out on a testbed which reproduces (on a small scale) a real prototype of an infrastructure network. Particularly, during the downstream from a desktop PC as a FH to a PDA as a MH, transmissions between these heterogeneous hosts may result in the degradation of an overall performance. Experimental analysis of the traffic profiles during the downstream has been done. In addition, we propose methods for improving the performance of such multimedia downstream. This paper consists of the following; we discuss the related works in Chapter 2. In Chapter 3, we show experimental analysis of TCP downstream and then present our proposed methods to improve performance. The experimental results are shown in Chapter 4. Finally, we conclude our works in Chapter 5. 2 Related Works Data transmission protocol adopted in this paper is TCP. TCP uses what it calls the congestion window to determine how many packets can be sent at one time. The larger the congestion window size becomes, the higher the throughput becomes [4]. Typically, the congestion window size in a wired network remains constant after short delay. However, it over a WLAN oscillates too rapidly. If the congestion window size increases rapidly, it can add to network traffic before the network has completely recovered from congestion. If congestion is experienced again, the congestion window size will shrink rapidly. This alternating increase and decrease in congestion window size causes the performance of data transmission over a WLAN to reduce remarkably [2]. Besides, we consider a manipulation of the send buffer and the receive buffer at the transport layer as well as an application buffer at the application layer to enhance the performance of a PDA by tuning TCP [7]. According to operating systems, the socket buffer size is different [7]. For each socket, there is a default value for the buffer size, which can be changed by the program using a system library call just before opening the socket. There is also a kernel enforced maximum buffer size. The buffer size can be adjusted for both the send and receive ends of the socket [4]. It varies with the operating systems. FreeBSD gives bytes of default TCP socket buffer size, whereas Windows 2000 and XP give 8192 bytes. To achieve maximal throughput it is critical to use the optimal sizes of the TCP send and receive socket buffer for the link we are using. If the buffers are too small, the TCP congestion window will never fully open up. If the buffers are too large, the sender can overrun the receiver, and the TCP window will shut down [4]. It is important to improve the TCP performance in wireless networks without any modifications of TCP syntax. The Snoop protocol of Balakrishan et al. [2] modified network-layer software at a base station to improved TCP performance in wireless

3 An Improvement of TCP Downstream Between Heterogeneous Terminals 741 networks. It used a split mode that a base station connected a fixed host and a mobile host. In split mode, a base station has to reveal data before it reaches the destination, thus violating the end-to-end semantics of the original TCP. 3 Experimental Analysis of TCP Downstream 3.1 A Test-Bed Infrastructure Network There are several simulation and analytical studies on a wired and wireless network, whereas in this work, we test a real system to measure the performance of multimedia transmission. Thus, we have designed and implemented VMS (Voice Messenger Systems). VMS is an infrastructure network that integrates a wired LAN based on Ethernet with a WLAN based on the IEEE standard. In our test-bed network, BS (Base Station) is simply AP (Access Point). A desktop PC and a PDA represents FH and MH, respectively. The hardware specification of the hosts used in the VMS is listed in Table 1. VMS is a kind of file transfer system. It is similar to short message services (SMS) available on digital mobile phones. Let us briefly explain how it works: a VMS client records one's voice and then sends it to the VMS server after converting it into a wave file format. The server receives this voice message and stores it in its hard disk. It transfers the message to the authenticated client that requests it. Table 1. The hardware specification of the hosts used in the VMS Host type CPU RAM NIC MH FH PDA PC Samsung S3C2440 (400MHz) Pentium 4 (2.4~3.0GHz) 128MB 1GB PCMCIA (11Mbps) PCI (100Mbps) Before we discuss about performance metrics, we should define the terminology to be used: Upstream is a process of transmitting data from a PDA as a MH to its server as a FH. On the contrary, we call downstream a process of receiving data for a PDA from its server. It moves in an opposite direction to the upstream. 3.2 An Analysis of Three Types of Transmissions We have analyzed three types of transmissions: upstream, downstream and wirelessto-wireless (PDA-to-PDA). Notice that when measuring the performance of such transmission, we located MH, i.e. PDA, within a 5-meter radius of an AP to maintain both good signal quality and strength. In the first traffic profile, we use a file size dimension up to 4688 Kbytes. These files are generated by recording one s voice for 60 to 600 seconds. In Fig. 1(a) the behavior of the transmission time with respect to the file size is shown when the packet size is 1460 bytes. The elapsed time to complete the downstream is slower than that to complete the upstream by 20% at maximum. Such difference becomes

4 742 Y.-H. Kim et al. greater as the file size becomes larger. In addition, the transmission time of the PDAto-PDA transmission is approximately equal to the sum of these two figures. In the second traffic profile, we use an application buffer size dimension of the sender up to 7168 bytes. In Fig. 1(b), the throughputs for the three types of transmissions are shown when we use both the file size of 2344 Kbytes and the packet size of 1500 bytes. One of the particular interests is the case of more than 1500 bytes of data, where it exceeds MTU (maximum transfer unit). In this test, fragmentation can take place at the sending host by varying the buffer size of the sender. If the buffer size is less than the MTU, the throughput in the downstream is superior to the upstream as depicted in Fig. 1(b). ª ª ª ª (a) The transmission times with respect to file size (b) The throughputs with respect to buffer size Fig. 1. The traffic profiles for three types of transmissions In that case, the possibility of fragmentation is very low and thus the transmission rate is moderate due to the relatively low traffic. We think that as the size of the application buffer at the sender increases it heightens the possibility of burst traffic and thus causes a high traffic. If the size of the buffer is larger, the throughput in the upstream is better. For the downstream, the degradation of the performance is due to the substantial differences in their computational complexities between these two heterogeneous hosts, a desktop PC and a PDA. Fig. 2. The throughputs with respect to both file size and buffer size

5 An Improvement of TCP Downstream Between Heterogeneous Terminals 743 In Fig. 2, we show the comparative analysis on 4 different file sizes during the downstream. As expected, the throughput increases as both the file size and the size of the buffer size at the sender increase. To see more details above, the traces for both the upstream and the downstream have been done by using the Network Analyzer [9]. Each arrival time of the first 11 packets during the downstream and the upstream is depicted in Fig 3. During the upstream, the VMS server as a FH receives 7 packets within 4.3ms. However, during the downstream a PDA as a VMS client receives only 6 packets within 7.1ms. Fig. 3. The arrival time of the first 11 packets during the downstream and the upstream (with the buffer size of 4096 bytes) 3.3 A Proposed Method of Improving the TCP Downstream From the analysis of the traffic profiles during the downstream, two aspects are clearly depicted: During the downstream (i) the window update is occurred frequently and (ii) the delayed ACK is generated very often compared to the upstream. This phenomenon causes the performance to reduce remarkably. First, the size of the window offered by the receiver can usually be controlled by the receiving process. This can affect the TCP performance. The size of the receive buffer is the maximum size of the advertised window for that connection. Thus, during the downstream we can use different receive buffer size dimensions. To be specific, buffers in routers are designed to be larger than the bandwidth-delay product [8]. To compute it, the bandwidth and the round trip time is measured by bps and 19ms, respectively, in the case of both the file size of 2344 Kbytes and the packet size of 1500 bytes. Thus, the lower limit is 810 bytes (approx.). Smaller buffers mean a higher packet drop rate [8]. Notice that the maximum allowable TCP window advertisement is bytes. However, too large buffer may lead to unacceptable delay of file access due to the smaller computational power of a PDA. So we should measure it to consider its impact. Second, in an application layer, an inter packet delay (IPD) refers to the time interval between the transmission of two successive packets by any host. In general, the transmission time increases as IPD decreases. However, if there is a clear difference in processing capability between sender and receiver, IPD should be

6 744 Y.-H. Kim et al. adjusted to give enough time to complete its internal processing for the low-end devices. The proposed method to be taken here is that longer IPD means the lower packet drop rate. In our test, we empirically chose IPD. Theoretically, IPD is chosen judiciously by considering network parameters, such as round trip time, the size of the transmit buffer and the receive buffer, and the capacity of the link, to avoid the degradation of the overall performance. 4 Performance Analysis of TCP Downstream 4.1 The Size of the Receive Buffer In Fig. 4 the behavior of the transmission time with respect to the size of the receiver buffer during the downstream is depicted. This diagram demonstrates that the large buffer size increases the advertised window for the connection. Fig. 4. The transmission time with respect to the size of the receive buffer An approximate 32% increase in throughput is seen by just increasing the buffer to bytes from 512 bytes. In our test, the lower bound on the buffer size is approximately 810 bytes as discussed in the Section 3.3. By just increasing the buffer to 1024 bytes from 512 bytes, the remarkable increase in throughput is obtained. Moreover, the diagram also demonstrates the un-correlation between the transmission time and the file access time. The elapsed time to access its internal files which resides in the memory of PDA can be kept constant with varying sizes of the receive buffer. 4.2 The Time Interval of the Inter-packet Delay In Fig. 5, the throughputs for two types of settings are measured with varying IPD: one is that IDP is set only at the sending side and the other is that IDP is set both at the sending side and the receiving side. In Fig. 5(a), during the upstream the

7 An Improvement of TCP Downstream Between Heterogeneous Terminals 745 behaviour of the transmission time with respect to the IDP is depicted. As expected, with no such delay the transmission time is the fastest independent of the transmission type. (a) The transmission times with respect to inter-packet delay during the upstream (b) The transmission times with respect to inter-packet delay during the downstream Fig. 5. The throughputs for both upstream and downstream with respect to IPD In the case of the downstream as depicted in Fig. 5(b), by setting the IPD of 5 ms only at the sending side achieves a best performance which improves by 7% compared to that without IPD. From the analysis of these empirical results the possible range of the IPD will be 1ms < IPD < 10ms. In addition, during the downstream we analyzed the packet losses with respect to IPD by using the AiroPeek [10]. Typically, all kinds of packet losses decrease with increases in IDP. In other words, reduces in packet loss cause the throughput to increase. 5 Conclusions We measured a performance of multimedia transmission between a PDA as a mobile host and a desktop PC as a fixed host in an infrastructure network. To evaluate such a performance more precisely, a test bed system called VMS that adopted TCP as a transmission protocol was built. The elapsed time to complete the downstream is slower than that to complete the upstream by 20% at maximum. During the downstream, the degradation of the performance is due to the substantial differences in their computational capabilities between these two heterogeneous hosts, a desktop PC and a PDA. From the analysis of the traffic profiles, two aspects are clearly depicted: During the downstream (i) the window update is occurred frequently and (ii) the delayed ACK is generated very often compared to the upstream. These cause the performance to reduce remarkably. First, by increasing the size of the receive buffer for a PDA the congestion window size of TCP becomes more stable. From our experiments, an approximate 32% increase in throughput is seen by just increasing the buffer to bytes from 512 bytes. In our test, the theoretical lower bound on the buffer size is approximately 810

8 746 Y.-H. Kim et al. bytes. Moreover, the results also demonstrate that the elapsed time to access its internal files which resides in the memory of PDA is kept constant with varying sizes for the receive buffer. Second, the inter-packet delay (IPD) should be needed to give enough time to complete its internal processing for the low-end device (PDA) during the downstream. By setting the inter-packet delay of 5 ms only at the sender achieves a best performance that improves by 7% compared to that with no such delay. From the analysis of the empirical results the possible range of the IPD will be 1ms < IPD < 10ms. References 1. Stevens, W. R.: TCP/IP Illustrated Volume 1: The Protocols, Addison-Wesley (1994) 2. Balakrishnan, H., Seshan, S., Amir, E., Katz, H.: Improving TCP/IP Performance over Wireless Networks, ACM MOBICOM (1995) 3. Nguyen, G. T., Katz, R. H., Noble, B., Satyanarayanan, M.: A Trace-Based Approach for Modeling Wireless Channel Behavior, In Proceedings of the Winter Simulation Conference (1996) Tierney, B. L.: TCP tuning guide for distributed application on wide area networks, ;login: The magazine of USENIX & SAGE, Vol. 26, No. 1 (2001) Zheng, B. and Atiquzzaman, M.: A Novel Scheme for Streaming Multimedia to Personal Wireless Handheld Devices, IEEE Transcations on Consumer Electronics, Vol. 49, No. 1 (2003) 6. G. Isannello, A. Pescape, G. Ventre, and L. Vollero, Experimental Analysis of Heterogeneous Wireless Networks, WWIC 2004 (2004) Karadia, D.: Understanding Tuning TCP, Sun Microsystems, Inc., (2004) 8. Lachlan L. H. Andrew, et al, Buffer Sizing for Non-homogeneous TCP Sources, IEEE Communication Letters, Vol. 9, No. 6 (2005) Analyzer web site, Wildpackets web site,