Simulation Analysis of TCP Perforriiance on IEEE Wireless LAN

Transcription

1 Simulation Analysis of TCP Perforriiance on IEEE Wireless LAN Yong Peng Haitao Wu Keping Long Shiduan Cheng National Laboratory of Switching Technology & Telecommunication Networks Beijing University of Posts and Telecommunications (BUPT), Beijing, China (100876) Phone: Abstract The TCP protocol is the most frequently used transport layer protocol in the Internet, and it is originally designed, improved and tuned to work efficiently on wired network where the packet loss is vely small. When packet loss is detected, TCP will trigger the congestion recovery algorithms. While in the wireless environment, the bit error rate of a wireless link is much higher arid a wireless connection might be temporally broken not only for the reason of congestion but also due to teniporal link impairment such as shadowing effect or buffer overjlow. So, it is valuable to research the current versions of TCP pe$ormance over a wireless environment. Wireless local area networks (WDW) provide users with high-speed wireless data access to the network and voice capabilities for telephone conversations. In this paper; we focus on the simulation analysis of the TCP performance over the DCF access mode of IEEE W M. 1. Introduction and Motivation With explosive growths in wireless services. and their subscribers, it is natural that supporting user mobility in the Internet is a research focus. The TCP protocol is the most frequently used transport layer protocol in the Internet, so it s valuable to research the current versions of TCP performance over a Wireless environment. But the traditional TCP is designed, improved and tuned to work efficiently on wired network where the packet loss is very small. In the wired network, when TCP detected the packet loss, it will consider that congestion has occurred and trigger the congestion recovery mechanism. While in the wireless environment, it is very possible that packet loss is caused by the high packet loss rate of the down layer. So how the high loss rate of wireless link will affect the TCP performance and how to improve the wireless MAC protocol to enhance the TCP performance remain to be researched. Wireless local area networks (WLAN) provide high bandwidth to users in a limited geographical area. The IEEE is in the process of developing an international WLAN standard identified as IEEE [1]. The IEEE is a standard for a WLAN covering both physical and MAC layers. It describes mandatory support for a I Mbps WLAN with optional support for a 2 Mbps data transmission rate. In this paper, we use the 2 Mbps transmission rate as default. At present, little work has been done to model the queue management mechanisms on TCP performance of IEEE WLAN, and the performance of bi-directional traffic flow (e.g. TCP flow) on the two DCF access modes (RTS/CTS enabled mode and RTS/CTS disabled mode) are also rarely researched, which is our research focus in this paper. To achieve the above research goals, it is necessary to investigate the impact of the critical MAC layer characters on the upper layer. So, we do two scenarios of simulations in this paper. We first use the simple model of two nodes to investigate some key characters of TCP performance over a Wireless LAN. We just use two nodes because we want to get the wireless unreliable aspect of the WLAN s impact on the upper layer, not the access aspect. The purpose of the simulations is to investigate the behavior of the different version TCP on the unreliable WLAN. We also use the buffer management algorithms on the queue (e.g. RED). After that, we do a more complex simulation of 2*N nodes, that is, N communication pairs to test the MAC layer access performance of IEEE DCF mode. Here we assume an error free radio channel because our focus here is the access aspect of WAN S impact on the upper layer protocol. The paper is organized as follows. Section 2 describes some background knowledge of the simulation analysis. Section 3 describes the simulation environment and results of the two scenarios. Section 4 concludes the paper. This work is sponsored by NOHA CHINA R&D CENTER. 5-7 x03= I /S I 0, EEE, 520

2 2. Background 2.1 WEE MAC Protocol The IEEE is a standard for a WLAN covering both physical and MAC layers. It appears to the layers above the MAC layer like any other IEEE 802.X LAN (e.g. Ethernet or Token Ring). Mobility aspects are handled at the MAC level or below. An IEEE WLAN generally consist of Basic Service Sets (BSSs) which are interconnected by a Distribution System (DS) to form an Extended Service Set (ESS) as shown below in figure 1. polls stations giving them the opportunity to transmit. PCF is optional, and its implementation is based on the implementation of DCF. In our simulation, we use the simulation model of ad hoc network. An ad hoc network is a deliberate grouping of stations into a single BSS for the purpose of internetworked communications without the aid of an infrastructure network. The former name of an ad hoc network in the IEEE standard is independent BSS (IBSS). In an ad hoc network, any station can establish a direct communications session with any other station in the BSS, without the requirement of channeling all traffic through a centralized access point (AP). Ad hoc network use the DCF (Distributed Coordination Function) access method support asynchronous data transfer on a best effort basis. 2.2 The Recovery Mechanisms of Different Versions of TCP Figure 1. IEEE Architecture The IEEE MAC protocol offers two types of service to its users: asynchronous and synchronous (or, rather, contention free). These types of services can be provided on top of a variety of physical layers and for different data rates. The asynchronous type of service is always available whereas the contention free is optional. The asynclironous type of service is provided by the Distributed Coordination Function (DO which implements the basis access method of the IEEE MAC protocol. We also call DCF a Carrier Sense Multiple Access with Collision Avoidance (CSMNCA) protocol. When packet is long, it is recommended that the DCF use an RTSKTS handshake before data transmission. RTS and CTS control frames can be used by a station to reserve channel bandwidth prior to the transmission of an MPDU so that bandwidth wasted due to collisions is minimized. RTS and CTS control frames are relatively small (RTS is 20 octets and CTS is 14 octets) when compared to the maximum size data frame (2346 octets). The synchronous type of service is Point Coordination Function (PCF) which mainly implements a polling access method. The PCF uses a Point Coordinator, usually the Access Point, which cyclically TCP is the most frequently used transport layer protocol. To gain better performance in all kinds of environments, people make enhancement on TCP and form many version of TCP. Reno, Newreno arid SACK are different versions of TCP. The main difference between the 3 versions of TCP is the behavior when multiple packets are dropped, which is always occurred in the wireless environment. In TCP Reno, people modified the fast retransmit operation to include Fast Recovery. The new algorithm prevents the communication path from going empty after Fast Retransmit, thereby avoiding the need to Slow-Start to re-fill it after a single packet loss. Reno significantly improves upon the behavior of Tahoe TCP when a single packet is dropped from a window of data, but can suffer from performance problems when multiple packets are dropped from a window of data. The Newreno TCP includes a small change to Reno algorithm at the sender that eliminates Keno s wait for a retransmit timer when multiple packets are lost from a window. In Newreno, partial ACKS received using Fast Recovery are treated as an indication that the packet immediately following the acknowledged packet in the sequence space has been lost, and should be retransmitted. Thus when multiple packets are lost from a single window of data, Newreno can recover without a retransmission timeout. The SACK TCP uses the TCP Extensions for high performance. The congestion control algorithms 521

3 l<eno. implemented in the SACK TCP are a conservative extension of Reno's congestion control. SACK differs from Reno when multiple packets dropped from a window of data. During fast recovery, SACK maintains a variable called pipe that represents the estimated number of packets outstanding in the path. The sender only sends new or retransmitted data when the estimated number of packets in the path is less than congestion window. [3]. 3. Siniulation Results of two nodes topology 3.1 Simulation Goals This report presents an in-depth simulation analysis of the TCP perfonnance on DCF access mode of BEE I MAC protocol. The first simulation scenario is composed of 2 nodes, that is, a single communication pair. We use this scenario to research how the TCP behaviors with the increase of packet discard rate and buffer size. For the wireless link, unlike the wired link, has unstable link character, so the buffer management algorithm will perform differently from the wired environment. The second simulation scenario is composed of 2*N nodes, that is, N communication pairs. We use this scenario to show the effect of RTS/CTS when packet collision is frequent. Although data integrity is a key requirement for data transmission, we assume an error free radio channel in this simulation scenario despite the fact that we are dealing with a very unreliable environment. This is because we are focusing on the multiple access aspects of the IEEE MAC protocol in this simulation part. the two nodes is 200 meters. Packet length is set to 1500 bytes. node 0 node 1 Figure 2. Scenario 1: Two Nodes Scenario Now first, let's have a look on the relationship between the packet discard rate and the throughput in case of no RTSKTS. We can see that even packet discard rate reached 20%, the difference between the 3 TCP versions remain the same. In figure 5, we can see that the congestion window don't drop, even when packet loss rate equals to 20%. So we can conclude that the MAC layer protocol has hidden the high packet loss rate of the wireless layer and the TCP protocol doesn't feel the down layer packet loss. We can also see from figure 3 that the increase of packet discard rate lead to the decrease of throughput, which is caused by the more frequent retransmission of the MAC layer data packets when packet loss rate increases. 3.2 Sirnulation Results of two nodes topology TCP Performance of Two Nodes Model In order to study the TCP performance over a WLAN, we first use the simple model of two nodes. FTP START TLME I 10.0s I FTP STOP TIME 1 so os I TCPWINDOWMAX I 40 1 i TESTED TCP VEKSIONS ~ Newreno. SACK ~ I DISTANCE I 200 m I I PACKETLENGTH bytes I I BUFFER SIZE I 50 (oacket) I I MACACCESSMODE I NORTSKTSDCF I Table 1: Two Nodes Topology Simulation Parameters In order to be more similar to the real communication environment, we use the shadowing model in the following simulation. The distance between Figure 3. Relationship Between The Pncket Discard Rate and Tluoughput Another result of the packet loss is the change of queue length. When packet loss is relatively low, packets usually can be transmitted without retransmission, so queue length is relative low. When packet loss is relatively high, the MAC packet retransmission will prolong the delay of packet. The long RTT (Round Trip Time) will cause the congestion window to increase slowly. So the average queue length will also have the trend of becoming short. In figure 4, when packet discard rate equals to IO%, average queue length reaches ;I climax. In figure 5, we can see the increase of congestion window in different packet discard rate. The more the packet is dropped, the slower the congestion window increases. 522

4 n NewReno 4 sack Figure 4. The effect of packet discard rate on queue length 008- B f t 0.021,,,.,,,,,, IO Buffer Size(packe1) Figure 6 The Effect of Buffer Size on Queue Delay Implementing RED on The Queue As we have seen above, the buffer length parameter affects the performance of Reno TCP much. So we apply some queue management algorithms on the queue and do some simulations in order to see whether those algorithms can improve the performance of WLAN or not. Figure 5. The Reno TCP Congestion Window Shape of Different Packet Discard Rate Scenatio Next, we do some simulations on the effect of buffer size on delay and throughput. With the increase of buffer size, different TCP versions appear little difference when queue delay increase, while appear relatively large difference on throughput. From figure 7 we can see that when buffer size is below or equal to 30, the Newreno and SACK TCP versions are far better than Reno TCP. And the difference between Newreno and SACK TCP is very small. The main difference between the SACK, Reno and Newreno is the behavior when multiple packets are dropped from one window of data. The reason that Reno has bad performance when buffer is small is that multiply packets are dropped in one window data, which is caused by buffer overflow. That causes the congestion window of Reno drop rapidly. 10 Buffer Size(packe1) Figure 7 The Effect of Buffer Size on Throughput 1 MAXDROPPROBARLLITY I I TESTED TCP VERSlON 1 Reno 1 MAC ACCESS MODE WP START TIME ftp STOP TIME PACKET LENGTH NO RTSKTS DCF 1460 bvtes Table 2. Simulation Parameters of Implementing RED on The Queue 523

5 4,10,20,30,40,50,70 ~~~ ~~ 1 From figure 8 we can see that the queue management algorithms don't affect the throughput much. The discard of buffered packet will cause the retransmission of TCP layer. That will lead to the decrease of throughput. While in figure 9 we can see that parameters properly tuned RED will greatly improve the queue delay character.._ o riio 1 RED drop RED drop head below: I NODENUMBtIZ ~ & FTP START TIME loo\ FTTPSTOPTEME a TESTHI TCP VERSION Newieno PACKET LENGTH MAC ACCESS MODE 1 Table 3. Simulation Parameters of Multi-Node Scenario 0.0 Node 0 Node 1 Ouaile iypc Figure 8 Effect of Queue Management on Throughput l'igure 9 Effect of Queue Management on Queue Delay 3.3 Simulation Results of Multi-Node As we know, the node number is an important parameter of a WLAN. With the increase of node number, the collision becomes more frequently. In order to investigate the performance of TCP on wireless LAN in multi-node environment, we now do the following simulation of multi-node. In this scenario, the even number stations are at one spot, while the odd number stations are at another point. The (2*N)th node will setup a FTP connection to the (2*N+l)th node. Because we are focusing on the multiple access aspects of the IEEE MAC protocol in this simulation part, we assume an error free radio channel in this simulation scenario despite the fact that we are dealing with a very unreliable environment. Simulation parameters are listed in table 3 1 Node 2bl-2 Node 2N-L Figure 10 Scenario 2: 2*N Nodes Scenario Throughput and Delay First, let's have a look at the effect of station number on throughput and average delay at the two scenarios of DCF (RTS) and DCF (NO RTS). First of all. the two curves have the same trend of decrease when node number increases. This is caused by the harder contention of MAC layer when node number increases. Second, we can see that when node number exceeds 10, the collision becomes more frequently, and the affect of RTS/CTS begins to be obvious. So we can conclude from the following figure that when packet is far larger lhdn the control frame, RTS/CTS can efficiently minimize the bandwidth wasted due to collision. O j DCF (NO RTS) 0.7!,,,,,,,,,,, I Node Number Figure 11 Throughput Comparison of The 2 MAC 524

6 Algorithms (10 seeds) number is large noRTSMACOelay &RTSMACOelsy ,.,,,.,.,.,.,.,., Nods Nmbrr Figure 13. NO RTS & RTS DCF Fairness (10 seeds) Figure 12. DCF (RTS) & DCF (no RTS) Average Delay Comparison Figure 12 shows the average delay comparison of DCF when RTS/CTS is used and RTS/CTS is disabled. First, we can see that with the increase of mobile nodes, the MAC layer delay increased. That is caused by contention and retransmission. From figure 12, we can also see that the queue delay of RTSKTS enabled DCF is less than the delay of RTS/CTS disabled DCF on the whole. Another conclusion is that the MAC delay of RTS/CTS disabled DCF is slightly longer than that of KTS/CTS enabled DCF. We can conclude that the RTS/CTS mechanism also improved the delay character of Wireless LAN Fairness The RTS algorithm will probably affect the fairness of the different communication pair. So it is necessary that we do some simulation on the fairness of different communication pairs. We use the fairness equation as follows [8]: f denote fairness, f(i) denote the ith communication pair. The following figure shows the comparison of the DCF fairness in case of RTS and NO RTS. The line links the mean value of 10 different seeds simulation results. We can conclude from the figure that with the increase of node number, the fairness of all the communication pairs decreases. We also see that the fairness of RTSlCTS enabled DCF is slightly better than that of RTS/CTS disabled DCF on the whole, especially when node 4. Summary and Conclusions In this paper we first research the case of two nodes scenario. From the simulation result, we can conclude that the IEEE MAC protocol has hidden the packet discard of the lower layer, while the buffer overflow will affect the TCP performance much. In the simulation of RED, we can see that because of the MAC retransmission, the throughput improves very little, while the queue delay improves much. Then we analyzed the simulation results of multi-node. From the result, we can see that the throughput and the delay and throughput characters of RTS/CTS enabled DCF are better than the RTS/CTS disabled DCF. RTS/CTS take effect and minimize the bandwidth wasted by collision when collision of large data packet is frequent. 5. References [I] IEEE , Wireless LAN Medium Access Control (MAC) and Physical (PHY) hyer Specifications, August [2] G. Anastasi and L.Lenzini, QoS provided by the EEE wireless LAN to advanced data applications: a simulation annlysis. Wireless Networks [3] Kevin Fall and Sally Floyd, Simulation-based Comp;iiisons of Tahoe, Reno, and SACK TCP, Coinpurer Cuniimicatiufu Rwirw. vo1.26, no.3, pp.5-21, July 1996 [4] George Xylomenos and Gcorgc C. Polyzos, TCP and UDP Performance over a wireless LAN, feee INI OCOM 99 [5] R. P. Crow, I. Widjaja, I. G. Kim, and P. Sakai, Investigntion OF the IEEE Medium Access Control (MAC) Sublayer Functions, IEEE infocom Y7 [6] Benny Bing, Measured Performance of thc IEEE I Wil-clcss LAN, IEEE Coniputer Society IEEEp [7] Ali Zahedi and Kevin Pahlavan, Capacity of a Wireless I.AN with Voice and Data Services, IEEE Transuctions on Conmiunicnrioiis 48 7 Jul2000 [8] R. Jain. The Art of Coniputer System Perfom~unce Analysis, John Wilwy and Sons,