Base Station. Mobile Host

Transcription

1 Modied TCP for Improved End to End Performance on Multimedia Wireless Network P. Venkataram, S. Chetan Kumar, K. Nagasesha eddy, S. Vishwanath y Protocol Engineering Lab ECE Dept. IISc Bangalore , India Abstract Transmission Control Protocol (TCP), a widely used reliable transport protocol uses end-to-end ow, congestion and error control, mechanisms. However this protocol is tuned to perform well on traditional wired network where packet loss is mainly due to congestion. They suer signicant degradations on a wireless network, which is characterized with high bit error rate and temporary connection loss due to hand-o. TCP responds to all loss by invoking congestion control and avoidance algorithms. n the other hand multimedia applications need seamless ow of data. In this paper we try to optimize end-toend performance of TCP connections on an internetwork, with hosts on both wired and wireless network. We can broadly think of two approaches, modications on the end This work is funded partly by Samsung India Software peration y S. Chetan Kumar is with Wipro Technology Solutions, Bangalore. chetansk@wipinfo.soft.net. K. Nagasesha eddy is with Tektronix Engineering Development (India) Ltd., Bangalore. S. Vishvanath is with Future Software, Chenni hosts and modications at the base station. The former requires changes to existing protocol on the stationary hosts, whereas the latter is a better approach. ur central idea is to perform re-transmission at the basestation and make TCP inherently unaware of it. The base station is made to sni all the packets that are transmitted across it. And since it is a fact that next hop to base station on the wireless network side is the end host itself, there is no possibility of out of order packet delivery from base station to mobile host. The base station is made to retransmit the lost packet before a duplicate acknowledgement is generated. For the trac from mobile host to xed host the base station is made to send a duplicate acknowledgement for any out-of-order packet snied by the base station. The base station keeps a record of number of acknowledgements received for every N number of packet snied. This gives an idea of underlying media status, depending on which the aggressiveness of the retransmission policy at the base station is changed. 1

2 1 Introduction People and their machines should be able to access information and communicate with each other easily and securely, in any medium or combination of media viz. voice, data, image, video, or multimedia, any time, anywhere, in timely, cost-eective way. As computing devices have shrunk in size, weight and cost and the recent research in mobile computing and wireless networks strongly indicates the evolution of computer environments, in which computing resources can used more exibly, as users will be freed from being physically connected to the underlying network. Communication over wireless network is characterized by limited bandwidth, high bit error rate and intermittent connection due to hand-o and volatile nature of the wireless media. eliable transport protocols, like TCP have been tuned, by the passage of time, to wired environment. TCP perform very well over media, like wired, where the packet loss due to media are almost nil. In these scenario the packet losses can be only caused by buer over ow (congestion) at the outers/gateways, in order to avoid the possible congestion collapse, TCP reduces the transmission rate, there by reducing the load on the outers/gateways[1]. Now consider TCP connection over lossy wireless link. If the wireless channel goes to a temporary bad state, due to signal fading or the connection gets intermittent due to hand-o. This causes the TCP to reduse its window size before retransmitting the lost packet and invoke congestion avoidance or control algorithm and resetting it retransmission timer. This will result in under utilization of network bandwidth. This would not be a much problem if TCP could increase it window size immediately as the below channel moves to good state, and the granularity in TCP timers were to be much ner. 1.1 elated Work Many solution are proposed by many researchers in this area, we have, here, summarized few of them. Indirect TCP (I-TCP) protocol[2,3]: This was one of the early protocols, it involves in splitting TCP connection between the sender and receiver into two separate connections at the base station. ne connection is from sender to the base station and another is between base station and the receiver. The advantage of splitting the connection is to isolate the wireless network with that of the wired one, and run a transport protocol which would be more specic to wireless network. This method, however has a major disadvantage that the end-to-end schematics of TCP acknowledgement is violated. The base station will acknowledge for for a packet, say, which may have not yet reached the mobile host. Selective Acknowledgments[4]: The existing TCP uses cumulative acknowledgment scheme. This happens to be a major set back when operating on a lossy channel, as loss recovery becomes dicult due to lack of sucient information of the lost segment. Several research works have 2

3 shown that a TCP with selective acknowledgment can perform much better then standard TCP, in a lossy environment. Sack (selective acknowledgment) was added as an option to TCP by FC However the differences over the use of SACK, prevented it from being adopted. ecently few new proposals are being made into FC This however needs major changes in all the existing hosts and acceptance as standard is not possible as of now. Snoop Protocol[5]: In this protocol, a snoop agent is introduced at the base station. This agent monitors the packets that are transmitted across it. This will detect the lost packet with few duplicate acknowledgements or time out. Then retransmit the lost packet if it were to be cached at the base station and suppress this duplicate acknowledgement reaching the sender TCP. ur approach is derived from this. This has disadvantages like not addressing the issue when TCP sender is a mobile host correctly, in which case SACK is the solution provided[6]. Link Layer etransmission[7]: A straight forward approach would be to make the channel more reliable, or make the higher layers to see the lower layer as a reliable link with lesser bandwidth. This can be achieved with FEC (Forward Error Correction) and retransmission of lost packet in response to Automatic repeat equest (AQ). This scheme has a major set- Ethetnet Base Station Mobile Host Inter Network Ethetnet Base Station Correspondent Host Mobile Host Figure 1: Typical Wireless Data Network back of timer interaction between the TCP. Suppose a packet is lost in the network and link layer retransmits the lost packet, before this lost packet is reached the destination and acknowledged, the TCP timer may expire and cause it to retransmit the packet as well. This competing retransmission by the transport layer and link layer often cause degradation in the network utilization The rest of the paper is organized as follows. In section 2, we describe the proposed modied TCP. In section 3, we present the performance of the proposed TCP and section 4 concludes the article. 2 The Modied TCP Most of the Multimedia applications, with good compression technique will occupy most of the band width. Some of these application will have stringent BE requirement. These applications will have to use 3

4 TCP for reliable transmission.ur aim is to provide TCP service for these application on wireless data network with better goodput (goodput is dened as ratio of actual transfer size to total bytes transmitted in the path[8]). The aim is to be achieved with any modication of existing TCP on the xed hosts. We have a good control over the TCP running on the mobile host and base station. We are viewing at the wireless network as shown in the gure 1. Where the Mobile Host (MH) is the mobile host which will be associated with a Base Station (BS) (The word base station and the router are used interchangeable here). Base station is also the router that route all packets from and to the wireless network. Correspondent or Fixed Host (CH) is a hosts on the internet, this can be on an another wireless network as well. Henceforth Mobile Host is referred as MH, Base Station as BS and Correspondent Host as CH. ur approach is to add a TCP layer at the base station, such that the IP layer while routing all packets will pass on copy to the higher TCP layer. The TCP layer at the base station will virtually bind at all the standard ports. n receiving an IP packet from CH, all TCP packets are copied to higher layer as though it were packets addressed to this BS. At this instant we cache all the unacknowledged packets, in the base station. And these packets are retransmitted from the base station to the mobile host, if they are lost in the wireless network, thus eectively shielding the sending TCP the loss that may occur in the Wireless Network. Looking at the typical wireless data network as shown in the gure 1, it can be easily deduced that mobile host is at a distance of one hop from the base station. Taking this fact into consideration it can be easily concluded that no packets arrive out of order from the base station to mobile host and vice-versa. Now when data is transfered from correspondant host to mobile host, under the assumption of streamlined data ow from the wired network, upon receiving a duplicate acknowledgement from the mobile host the base station may not have to wait till third duplicate acknowledgement for retransmission. Now it may not be mandatory to suppress the duplicate acknowledgement, since the sender will have to wait till the third duplicate acknowledgement. Also the base station keeps a record of number of acknowledgements received for every N number of packet snied. This gives an idea of underlying media status, depending on which the aggressiveness of the retransmission policy at the base station is changed. This is achieved by reducing the T at the base station and forcing the base station to time out for retransmitted packets, to a virtual value ahead of actual time-out, and thus pumping the more packets to the network, when there is transient bad state in the underlaying MAC layer. Unlike snoop protocol the base station in our approach has a good knowledge of TCP and converse to I- TCP: Indirect TCP [2] we are not splitting th TCP at the base station which will cause much over head at the base station. The abstract model of network we are looking is as shown in the gure 2. 4

5 TCP IP Host A IP packets, sent to TCP Virtual path of TCP MTCP IP Base Station a copy is running here TCP IP Host B Figure 2: Abstract Model of MTCP protocol 2.1 Connection Establishment Phase Now we shall consider the proposed protocol operation. Consider there is a host CH, which is a host on the wired network, tries to connect to a host MH on the wireless network, which has BS as the base station. When CH sends a syn to MH, at the base station, while forwarding the packet to the next hop (MH in this case) will also make a copy for itself. The TCP running at the base station, virtually binds to all services, will accept this syn signal from CH, creates a new TCB (Transmission Control Block) for this connection and waits in LISTEN state. As the same request is accepted by the MH, will send syn,ack. This will cause the BS to change the state to SYN_CVD from LISTEN, upon receiving ack from the CH the BS will change it state to ESTAB- LISHED. Thus the TCP states of BS are changed based on the packets send and received from the MH or CH. 2.2 Data Flow from Correspondent Host Now that CH has initiated the connection and start sending data to MH. TCP at the CH implements sliding window scheme to transmit packets. The window size at any instant of time will be minimum of congestion window and receiver advertised window, (thus TCP achieves ow control and congestion control). TCP is byte stream protocol, where each byte is associated with a sequence number. TCP groups chunks of data bytes into a packet or segment. Each segment is uniquely identied by its begin and end sequence number. The M-TCP sitting at the basestation will keep track of these sequence numbers for particular connection. When a TCP packet is send from CH to MH, and this is new packet (the sequence number is more then the highest sequence number seen) M-TCP cache this packet. A timestamp is added to this packet, in order to estimate round-trip-time(rtt) of the wireless link. This rtt is used to time out for retransmission. Now if this new packet received happens to be out of sequence, probably due to congestion in the wireless network, M-TCP will still cache the packet, as the packets are forwarded to the MH. There are possibilities of receiving a duplicate packets, when a duplicate packet is received and it does not exists in the cache, this indicates a loss of acknowledgement in the wired network and sender timeout. We will discard this packet and send an ack for this packet. n the other hand a duplicate packet is received and it exist in the cache, this also indicate sender time out, but loss of packet in the wireless link, at this instant 5

6 we will retransmit the lost packet to the MH. As the data ows from CH to MH, ack in the opposite direction. TCP uses cumulative acknowledgemet technique. These ack free the cache buer. When a new ack is received and this acknowledges the packets received at the MH, we will clear these packets in the cache. Also the rtt for the wireless link is updated at this time. The rtt is updated only once in a window size and done so only when there is no retransmission[9]. Arrival of dup-ack (duplicate acknowledgement), indicates the packet loss in the wireless link, this is usually the case as the wireless link is more lossy and TCP generates a dup-ack for each segment that it receives after a segment is lost. Now there are two possibilities, the packet which caused dup-ack may be in the cache (a more normal case). The BS will retransmit this lost packet, it is also possible to maintain a separate queue at the link layer for the packets to be retransmitted to improve the performance. Unlike the standard TCP we may not wait till the third dup-ack, as we know that MH is at a distance of one hop from the BS. n the other hand there may be a dup-ack for a packet that is not cached. This is the case where the packet is lost in the wired link, may be due to congestion, no action is taken as the congestion control is well handled by the CH. Another process that indicates loss of packet in the wireless network is timer expiring at the BS. The M-TCP will maintain a smoothened rtt which is, TT = (1 - )*old_tt + * cur_rtt. T is a constant multipule of TT, where constant multipule is between 1 and 0, usually If the BS and M-TCP doesn't see acknowledgement with in this T, then the packet need to be retransmitted, as it would be lost, and rest of the subsequent packets are also lost thus no dup-ack. Now in a wireless environment these type of losses are more common, due to multipath fading or due to shadowing. Under these sutiations it becomes necessary to sence the underlying media for better performance. MTCP at the BS keep track of number of ack n, received for every N packets send accross. The ratio n:n, which we call the media factor, is a good indicator of the underlying media. (This relation can be easily deduced as we consider one hop distance from BS to MH). As the media factor becomes closer to zero, almost all packets transmitted are lost, and as the media factor approches one, (it can never be equal to one) the goodput of the network will be one. In our MTCP we make the T inversely proportional to media factor, depending on which the timer at the MTCP will expire and retransmission takes place. Thus if the wireless link goes to transient bad state we make a retransmission at much shorter time interval, and see that more packets are pumped into the network. 2.3 Data Transfer from Mobile Host We shall now consider data transfer from MH. As the packets are transmitted from MH to CH bulk of the packet loss occurs in the wireless link and BS in no way can come to know if the packet loss has occured in the wireless or congestion has occured some where ahead. Few proposals of SACK TCP for MH were proposed[4]. But we con- 6

7 strain changes only to BS, so SACK TCP at MH is ruled out. Now if a packet loss were to occur in the wireless link, would cause MH to invoke congestion control and avoidance process. The MH would have to wait for 3 dup-ack to conclude a packet is lost and would go for retansmission of lost packet, ofcourse after invoking congestion control and avoidance alogrithm[10]. Now again since the MH is at one hop distance from the BS, BS can know the lost packet if it receives out of order packets. At this instant of time, the BS will issue 3 dup-ack for the lost packet so that MH can retransmit the lost packet. We have to compromise here as we constrain all our changes to BS. This however will increase the performance as MH can come to know about the lost packet at the earlier stage. fcourse the MH will discard all ack sent by the CH for those packets earlier than the lost packet [11]. The BS will retransmit all those packets earlier to lost packet, upon receiving triple dup-ack from the CH. 2.4 Mobility Suport We have built up this protocol with underlying IP being Mobile IP, the mobility issue and hand-o is dealt with Mobile IP, this FC gives the full details[12]. At each basestation we have provided a service, from which every other base station can get TCP connection status upon request, based on its permanent IP number (IP number at its home network). When a MH moves from it home network to foregin network,f A 1, and gets a new IP number, the foregin agent can request the status of connection for the new MH with its home agent. As the MH moves to a new for- sequence jala.pclab:20_==>_protocol.ece.iisc.ernet.in:22176 jala.pclab:20_==>_protocol.ece.iisc.ernet.in:21530 number (time sequence graph) Z Z Z Z YN 0 0 s s s s s time Figure 3: Time Sequence plot of TCP and MTCP for data ow from CH to MH eign network,f A 2 and gets a new IP number and registers with the home agent, the home agent will request for the connection status from the old foreign agent,f A 1. The new foreign agent, F A 2 will request for the connection status from the home agent and home agent will respond with the updated connection status. It is to be noted that connection status also includes the current cache for that connection. 3 Performance of MTCP We have performed several experiments with Modied TCP on wireless testbed. Testbed consisted of P100 wireless host each with wavelan card. The base station is a p-133 with wavelan card and NE2000 compatible ethernet card. The wavelan card [13] supports a raw bandwidth of 2Mbps. With out any modication and in the prescence of no packet loss, a throughput of 1.6Mbps is achieved with TCP overhead. The throughput degraded drastically in the presence of packet loss. 7

8 With the modied TCP we could achieve same througtput in with the prescecne of 20% of packet loss.we also found comparatively better performance in the prescence of higher percentage of loss. In the gure 3. above we have shown time-sequence plot of data transfer from a xed CH on wired network to a MH, with existing TCP and with MTCP. The data transferd is an MPEG le of the order of 15MB. Tcpdump network monitoring utilitiy is used to collect the data for the plot. These plots are plotted with xplot utility. 4 Conclusion In this paper we have discussed the problems of reliable protocol like TCP when the under laying media is a hybrid media, where the loss characteristics dier signicantly. We have developed a protocol that improves the end-to-end performance of the TCP on heterogenous media. All changes that can be allowed in the system are conned to Base Station. ur approach is to add a TCP layer at the Base Station, with basic idea of caching packets and making local retransmission. The TCP layer at the Base Station will have limited TCP intelligence to make local retransmission at varied time out depending on the status of the underlaying media. Also the TCP at the BS will provide service to mobility agents to maintain the status of connection as the mobile host moves across the networks. eferences [1] V. Jacobson. Congestion Avoidance and Control. Proc. ACM SIGCMM '88, pp [2] A. Bakre and B.. Badrinath. I- TCP: Indirect TCP for Mobile Hosts In Proc. 15th International Conf, on Distributed Computing Systems (ICDC S), May [3]. Yavatkar and N. Bhagwat. Improving End-to-End Performance of TCP over Mobile Internetworks. In Mobile 94 Workshop on Computing Systems and Applications, December [4] Kevin Fall and Sally Floyd. Simulation-based Comparison of Tahoe, eno, and SACK TCP. ftp://ftp.ee.lbl.gov/papers/sacks_v1.ps.z [5] Hari Balakrishnan, Srinivasan Seshan, Elan Amir and andy H. Katz. Im proving TCP/IP Performance over Wireless Networks. Proceedings of the 1st ACM International Conf. on Mobile Computing and Networking (Mobicom), November [6] Hari Balakrishnan, Srinivasan Seshan, and andy H. Katz. Improving eliable Transport and Hando Performance in Cellular Wireless Networks. ACM Wireless Networks, 1(4), December [7] H. Chaskar, T. V. Lakshman, U Madhow. n The Design of Interfaces for TCP/IP ver Wireless. Milcom '96. [8] Hari Balakrishnan, Venkata N. Padmanabhan, Srinivasan Seshan and andy 8

9 H. Katz. A Comparison of Mechanisms for improving TCP Performance over Wireless Links. SIGCMM '96 8/96 CA, USA, pp , [9] P.Karn and C.Partridge. Improving ound-trip Time Estimates in eliable Transport Protocols. SIGCMM 87, August [10] W. Stevens. TCP Slow Start, Congestion Avoidance, Fast etransmit, and Fast ecovery Algorithms. FC 2001, January [11] Information Sciences Institute University of Southern California. Transmission Control Protocol DAPA Internet Program Protocol Specication, FC793 September 1981 [12] Network Working Group C. Perkins, Editor. IP Mobility Support, FC2002 ctober 1996 [13] David Eckhardt and Peter SteenKiste. Measurement and Analysis of the Error Characteristics of an In-Building Wireless Network. SIGCMM '96 8/98 CA, USA. pp ,