TCP: Transmission Control Protocol, is a connection-oriented (connection oriented), reliable, byte stream-based transport layer (Transport layer) communication protocol, the IETF's RFC 793 instructions (specified). In the simplified OSI model of computer networks, it completed the fourth layer functions specified by the transport layer, UDP is another important within the same layer transmission protocol.
Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 108 TCP Enhancement: Fast Active Queue Management Scalable Transmission Control Protocol TABASSAM NAWAZ, MUHAMMAD SALEEM MIAN, HAFIZ ADNAN HABIB Telecommunication & Information Engineering Department University of Engineering & Technology Taxila, Punjab PAKISTAN firstname.lastname@example.org, email@example.com, firstname.lastname@example.org Abstract: - FAST TCP is Fast Active queue management Scalable Transmission Control Protocol. FAST TCP is an alternative congestion control algorithm in TCP. It is designed for high speed data transfers over large distance. FAST TCP is assumed to be the “successor” of TCP. Lot of people already worked on this subject. Their research work and results yield a new dimension for the entertainment and scientist eras. This research paper is the comprehensive study on FAST TCP. Key-Words: - TCP, Fast TCP, Congestion Control, Fast TCP review, token based F-TCP. 1. Introduction: the speed of the previous one, and repeats Computer systems worldwide use TCP/IP the process, getting slower each time, until it protocols to communicate because TCP/IP succeeds. This means that even minor provides the highest degree of glitches on the line can make a connection interoperability, and runs over more network very sluggish. Because FAST TCP uses the technologies than any other protocol suite. same packet sizes as regular TCP, the The robustness of TCP is the main reason hardware that carries messages around the for its large-scale deployment. Besides, TCP net will still work. The difference is in is one of a few transport protocols that have software and hardware on the sending congestion control mechanisms. With computer, which continually measures the acknowledgments and time-out based time it takes for sent packets to arrive, and congestion control mechanism, the how long acknowledgements take to come performance of TCP is inherently related to back. This reveals the delays on the line, the bandwidth delay product of the giving early warnings of likely packet connection. For a TCP connection, losses. The FAST TCP software uses this to congestion might occur in those nodes (e.g., predict the highest data rate the connection routers, IP/ATM access nodes) along the can support without losing data . traversing paths, which mostly results in Dr. Jian Ma originally proposed F-TCP in packet loss . NRC, and it has become an ATM Forum TCP breaks down large files into small draft. Its basic idea is to avoid congestion in packets of about 1500 bytes, each carrying intermediate nodes by effectively controlling the address of the sender and the recipient. acknowledgement (ACK) flow, that is, delay The sending computer transmits a packet, the ACKs to inform the source that the waits for a signal from the recipient that network will be congested . acknowledges its safe arrival, and then sends the next packet. If no receipt comes back, 1.1 Basic principle of the F-TCP the sender transmits the same packet at half flow control: Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 109 The present simple FAST TCP (F-TCP) flow control relates to end-to-end flow 1.3 Comparison of Standard TCP control in packet network where with FAST TCP: Transmission Control Protocol (TCP) is We will take a look on some properties used as transport layer protocol. The most comparison between standard TCP and critical problem today in the Internet is the FAST TCP. long control time of the TCP flow control 1.3.1 Infrastructure problem which results buffer oscillation, low link TCP: Does not solve infrastructure problem utilization and low throughput. The main FAST: Does not solve infrastructure objectives of the F-TCP flow control are to problem remedy these problems by early informing 1.3.2 To get high throughput TCP source that the network will be TCP: The efficiency of the (congestion congested, and to direct the TCP source to control algorithm in the) current TCP slowdown its output rates. The basic idea of implementation drops steadily, and the the scheme is to delay the ACKs being protocol eventually becomes a performance transferred from the destination towards the bottleneck itself, as the network sender. This can be done at the same infrastructure scales up in capacity . network point where congestion has been FAST: FAST TCP aims to remove this detected, or, alternatively, a network point bottleneck: it is scalable to networks with detecting overload or congestion can direct large bandwidth-delay product . another network point to delay the ACKs. 1.3.3 Performance improvement in low speed networks 1.2 Simple FAST- TCP: TCP: If the bottleneck in the end-to-end The basic idea of the F-TCP flow control is path is the 10Mbps or 100Mbps Ethernet to delay the ACKs being transferred from card, we expect the current TCP the destination towards the source to inform implementation to be quite efficient, so there the source that the network will be is not much to improve . congested, and to direct the TCP source to FAST: If the performance of the current slowdown its output rates. This can be done TCP implementation is poor even at such at the same network point where congestion speeds, then FAST TCP may or may not has been detected or a network point provide significant improvement depending detecting overload can direct another on the reason for the poor performance . network point to delay the ACK. When 1.3.4 High speed networks and wireless detecting overload, F-TCP delays ACK on networks. the backward path instead of discarding TCP: The current TCP performs poorly in packets on the forward path to inform the two types of networks . TCP source . FAST: FAST TCP is optimized for the Router Congested former and believe it can be tailored to provide significant benefit in wireless networks as well . 1.3.5 Delay-based congestion control TCP: Delay-based congestion control has Destination been proposed since the late 80s by Jain and Source Delayed ACK many others, notably Brakmo and Peterson ACK in TCP Vegas. We believe its advantage over loss-based approach is small at low speed, but decisive at high speed . FAST: This does not mean that it is futile to use delay as a measure of congestion, but rather, that using a delay-based Fig 1: Prototype of Simple FAST TCP Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 110 algorithm to predict loss in the hope of Fig 1 shows the prototype of F-TCP helping a loss-based algorithm adjust its exploited in routers. The mechanism of window is the wrong approach to FTCP could be divided into three parts: address problems at large windows. congestion detection, ACK’s identification and delaying ACKs. A fixed threshold for Instead, a different approach that fully the forward buffer occupancy is set, so that exploits delay as a congestion measure, congestion is notified once the buffer augmented with loss information, is occupancy exceeds the threshold. ACKs needed . flows are delayed according to the state 1.3.6 Difficulties of the current TCP at (CONGESTION or NON-CONGESTION), large windows that is, when no congestion occurs, ACKs TCP: Four difficulties contribute to the leak by a normal rate, otherwise, by a poor performance of current TCP fraction of the normal rate. We set normal implementation in networks with large rate the same as the rate of data packet in the bandwidth-delay product:  forward path. The fraction is set to half that • At the packet level, linear increase the rate is halved when congestion is by one packet per Round-Trip Time detected. Besides, determining the rate of (RTT) is too slow, and Delaying ACKs is also a hazard problem. multiplicative decrease per loss For F-TCP, ACKs should be delayed event is too drastic. according to the network traffic conditions, • At the flow level, maintaining large while traffic in real network changes so average congestion windows quickly and frequently that it is difficult to requires an extremely small grasp. So a scheme is developed called equilibrium loss probability that is token-base F-TCP . hard to achieve in practice. Forward Buffer • At the packet level, oscillation is Data Server unavoidable because of the binary nature of the congestion signal (packet loss). • At the flow level, the dynamics is unstable, leading to severe Token based oscillations that can only be Mechanism reduced by the accurate estimation of packet loss Backward Buffer probability and a stable design of the flow dynamics . Server ACK FAST : Fig: 2 Token Based FAST TCP The prototype of the token-based FTCP is FAST TCP is equation-based, hence shown in Fig.3. When data packets arrive or avoiding packet level oscillation, leave and the resource condition changes, FAST TCP has stable flow this mechanism collects associated dynamics, information and recalculates the number of FAST TCP uses queueing delay, tokens. This mechanism has many potential rather than loss probability, as the advantages as following. Since this scheme main measure of congestion . is one type of F-TCP, it also possesses the merits of F-TCP, such as fully avoiding 1.4 Token based F-TCP: packet loss, shortening buffer capacity, smoothening traffic etc. It does not need to determine the threshold or the rate of Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 111 delaying ACKs. ACKs are constrained fashion, and in the presence of current TCP closely related to current network resource flows needs a lot more evaluation . (i.e., spare buffer occupancy) not the ns-2 characterization of data traffic. As a result, this scheme is significantly robust. This scheme is also very simple and easy to algorithm. In TCP, delayed ACKs allows data receivers to refrain from sending ACK for every incoming data packet. Although delayed ACKs can reduce the number of packets sent by the receivers, excessive delays on ACKs will disturb the round-trip timing and inherent self-clocking of TCP. 1.5 In ATM Networks: The essential idea of FAST TCP (F-TCP) is to delay TCP acknowledgment (ACK) traveling towards its TCP source through a node where its forward channel is The congestion control mechanism of TCP congested. It can be seen that: 1. F-TCP separated into four components in fig 4. smoothes the peak of the TCP flow, These four components are functionally consequently F-TCP reduce the requirement independent so that they can be designed of ATM buffer; separately and upgraded asynchronously . 2. With the same size of ATM buffer, F- TCP reduces the probability of overflow and Data Window Burstiness as a result, improve the TCP throughput; 3. F-TCP reduces ATM buffer oscillation, Control Control Control since in most time the TCP is in congestion avoidance phase after short period of slow start phase, the flow is fairly smooth and the ATM buffer utilization is improved . Estimation 2. METHODOLOGY 2.1 WAN in LAB at netlab.CALTECH.edu They have described the development of TCP Protocol Processing FAST TCP, from background theory to actual implementation and its first Fig 4: FAST TCP Architecture demonstration. Unlike TCP Reno and its They presented experimental results of their variants, FAST TCP is delay-based. This first Linux prototype and compared its allows it to achieve high utilization without performance with TCP Reno, HSTCP and having to fill the buffer and incur large STCP. They have evaluated these algorithms queuing delay, as loss-based algorithms not only in static experiments, but also often do. It achieves proportional fairness dynamic environments where flow comes and does not penalize flows with large and go; not only in terms of end to end RTTs. Whether FAST TCP can converge throughput, but also queue behavior in the rapidly, yet stably, to a fair allocation in a network . dynamic environment where flows of heavy- FAST tailed sizes join and depart in a random Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 112 5 5 x 10 x 10 10 10 9 9 8 8 7 7 throughput (Kbps) throughput (Kbps) 6 6 5 5 4 4 3 3 2 2 1 Steady throughput 1 0 0 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 sec sec LINUX FAST 5 10 x 10 2000 1500 (pkt) 9 1000 avg 8 q 500 7 0 0 0.5 1 1.5 2 throughput (Kbps) 6 5 4 x 10 x 10 5 5 cumulative loss (pkt) 4 4 3 3 2 1 2 0 0 0.5 1 1.5 2 1 4 x 10 80 throughput (pkt/ms) 0 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 sec 60 Dynamic sharing on Dummynet 40 • capacity = 800Mbps 20 • delay=120ms 0 0 0.5 1 sec 1.5 2 4 • 3 flows x 10 • iperf throughput LINUX 2000 • Linux 2.4.x (HSTCP: UCL) 1500 (pkt) 1000 HSTCP avg q 500 5 x 10 10 0 0 0.5 1 1.5 2 5 4 x 10 x 10 9 5 cumulative loss (pkt) 4 8 3 2 7 1 0 0 0.5 1 1.5 2 throughput (Kbps) 6 4 x 10 80 throughput (pkt/ms) 5 60 4 40 20 30min 3 0 0 0.5 1 1.5 2 sec x 10 4 2 Dynamic sharing on Dummynet • capacity = 800Mbps 1 0 0 1000 2000 3000 4000 sec 5000 6000 7000 8000 9000 • delay=120ms STCP • 14 flows • iperf throughput • Linux 2.4.x (HSTCP: UCL) Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 113 HSTCP 1 TCP Reno 2000 0.9 FAST TCP HighSpeed TCP 1500 Scalable TCP 0.8 (pkt) 1000 0.7 avg q 500 0.6 CDF 0 0 0.5 1 1.5 2 0.5 H x 10 5 4 x 10 S 5 0.4 T cumulative loss (pkt) 4 0.3 C 3 P 2 0.2 ~ 1 R 0.1 0 0 0.5 1 1.5 2 0 4 x 10 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 80 Jain’s index Fairness throughput (pkt/ms) 60 40 20 Stability 0 1 0 0.5 1 1.5 2 sec 4 x 10 0.9 STCP 0.8 2000 0.7 1500 (pkt) 0.6 1000 avg CDF 0.5 q 500 0 0.4 0 0.5 1 1.5 2 5 4 x 10 x 10 0.3 5 cumulative loss (pkt) 4 0.2 TCP Reno 3 FAST TCP 0.1 HighSpeed TCP 2 Scalable TCP 1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0 0.5 1 1.5 2 Stability 4 x 10 80 throughput (pkt/ms) 60 40 2.2 Token Based F-TCP 20 For the long control loop problems, a 0 0 0.5 1 1.5 2 mechanism named FAST-TCP is proposed sec to avoid congestion in the intermediate 4 x 10 Aggregate throughput nodes by effectively controlling ACKs flow 1 TCP Reno traverse the same node as its forward 0.9 FAST TCP HighSpeed TCP packets. The algorithm contains three 0.8 Scalable TCP parameters: token buffer capacity: BT; 0.7 ACKs buffer capacity: BA; number of B 0.6 tokens: NT. When data packets arrive or leave, the resource condition changes, this CDF 0.5 0.4 mechanism collects the information and 0.3 calculate number of tokens. Some temporary 0.2 variables are needed: average packet length: Lp; data packet counts: Pc. The procedure of 0.1 its implementation can be depicted with the 0 100 200 300 400 500 600 700 800 pseudo-code: Throughput (Mbps) Initialization: Packet count=0; BT = NT =Forward buffer capacity/Default packet Fairness length, then, If a data packet arrives at the forward buffer, 1. Recalculate the average packet length Lp and BT packet count = packet count+l ; Lp=(Lp+input packet length)/packet count BT =Forward buffer capacity/ Lp B Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 114 2. Calculate NT It can be clearly seen that after carefully NT =Spare Forward buffer capacity/Lp setting the trigger condition of delay time NT = min(BT, NT); End increase for F-TCP and backward ACK If a data packet leaves the forward buffer, delay time, ATM switch buffer will not - Calculate NT overflow even though the buffer of the ATM NT =Spare Forward buffer capacity/ Lp switch is relative small (500 cells). In this NT = min(BT, NT); End simple example, the overflow of the buffer If NT >0 an ACK is served, of the ATM switch can be completely NT =NT --l; End  avoided. The reason of this conclusion will be analyzed below. For 536 bytes of the 2.3 In ATM Networks: packet of data from TCP layer, the ATM The implementation of F-TCP in the third switch receives: 536 bytes of data + 20 bytes layer and add some new statistics in the of TCP header + 20 bytes of IP header + 8 ATM switches to monitor the utilization and bytes of LLC header + 8 bytes of AAL5 occupancy of the ATM buffer. They use a trailer = 592 bytes. These are padded to very simple example to show the produce 13 ATM cells. Since the receiver effectiveness of F-TCP. Both the client and waits no time after receiving a segment the server in the network use F-TCP. The before sending an ACK, each ACK will client downloads one big file from the acknowledge 536 bytes data in the TCP server. The following is some important layer. And for each ACK received by the parameters of the network . sender, at most twice mount of data will be 2.3.1 Client and Server sent out. In another words, at most 2* 13 = ATM buffer capacity: 500 cells 26 ATM cells will arrive at the buffer from IP Forwarding Kate: 10,000,000 packet/sec. the forward link and 2 ATM cells from TCP Initial RTO (Retransmition Timeout): backward link which are segmented from a 0.5 sec. ACK packet before the increasing of the TCP Maximum ACK Delay: 0.0 sec. (This delay time of backward ACKs by F-TCP parameter is the maximum time the TCP takes effect. When the buffer occupancy waits after receiving a segment before reaches as high as 500 - 26 - 2 = 472 cells, sending an acknowledgment.) additional specified delay time is triggered TCP Maximum KTO: 10 sec. to be added to the consecutive backward TCP Maximum Segment Size: 536 bytes ACKs. Therefore the consecutive ACKs stay TCP Minimum RTO: 0.25 sec. more time in the 1P layer of the sender node TCP Receive Buffer Capacity: 2,000,000 before they are passing to the TCP layer. By bytes carefully selecting the delay time, the buffer Trigger of Increase of Backward ACK occupancy returns below 472 cells as a Delay Time if F-TCP is Enabled: ATM result of the cells being transferred by ATM buffer occupancy is greater than or equal to switch . 472 cells 2.3.5 Results when F-TCP is disabled Increasement of Backward ACK Delay It can be seen that without F-TCP, the buffer Time: 7.634E-5 sec . of ATM switch faces to overflow. As a 2.3.2 ATM Switches (SW 1 and SW 2) result, the throughput is very low. With the ATM Buffer Capacity: 1000 cells same network configuration and the ATM Switch Fabric Delay: 0.0 sec . parameters of the network elements, when 2.3.3 Links F-TCP is enabled, more than 2M bytes of Data Kate: 155,520,000 bit/sec. data are sent , while when FTCP is disabled, Delay: from Client to SW I is 5E-6 sec., only less than 400K bytes of data are sent in from SW 1 to SW 2 is 0.12 sec., from SW 2 the same period (10 sec.) One time of ATM to Server is 5E-6 sec . buffer overflow will result in the loss of one 2.3.4 Results when F-TCP is enabled TCP segment and consequently reducing the Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 115 TCP throughput significantly in plain TCP remarkable thing about FAST TCP is that it . uses the existing Internet. The secret is in software at the sending point that parses the 3. FAST TCP: Benefits, data into network-compatible packets that avoid typical Internet congestion as they Achievements and Future weave their way to their ultimate destination Planning: . 3.1 Advantages & Disadvantages of FAST TCP: 3.3 Use standard packet size: Advantages of FAST TCP The protocol is called FAST, standing for • FAST TCP is just like TCP in the Fast Active queue management Scalable sense that any application, such as Transmission Control Protocol (TCP). The FTP, using TCP will use FAST TCP FAST protocol sustained the speed using once the patch is installed. You standard packet size, stably over an don’t need any special programs to extended period on shared networks in the use FAST TCP. presence of background traffic, making it • In principle FAST TCP can be adaptable for deployment on the world's transported over IPV6 or IPV4. high-speed production networks. The ability Unfortunately the Linux to demonstrate efficient high performance implementation splits the TCP throughput using commercial off the shelf source into an IPV6 and IPV4 part hardware and applications, standard Internet and we only have an IPV4 packet sizes supported throughput today's implementation presently. An IPV6 networks, and requiring modifications to the implementation is on the roadmap ubiquitous TCP protocol only at the data . sender, is an important achievement. The Disadvantages of FAST TCP problem today is that this algorithm cannot • Reverse path congestion reduces scale to anticipated future needs, when the throughput. networks will be compelled to carry millions of uncompressed voice calls on a single path • Many FAST sources cause buffer or support major science experiments that overflow and packet loss. require the on-demand rapid transport of • Route change may reduce gigabyte to terabyte data sets drawn from throughput . multi-petabyte data stores. This protocol problem has prompted several interim 3.2 Avoid typical Internet remedies, such as using nonstandard packet congestion: sizes or aggressive algorithms that can The transmission control protocol (TCP) is monopolize network resources to the seen as the dominant transport protocol. The detriment of other users. Despite years of current stability of the Internet depends on effort, these measures have proved to be the end-to-end congestion control of TCP. ineffective or difficult to deploy. Using TCP does not perform well in high-speed standard packet size that is supported wide area networks. To achieve a steady- throughout today's networks, the current state throughput of 7.2Gbps with 1500 byte TCP typically achieves an average packets and a 100 ms round trip time (RTT), throughput of 266 Mbps, averaged over an for example, the packet loss rate must be hour, with a single TCP/IP flow between less than 4.17 × 10−10. This is beyond the Sunnyvale near SLAC and CERN in limits of achievable fiber error rates. In Geneva, over a distance of 10,037 addition, TCP requires 40,000 RTTs, or kilometers . almost 70 minutes, to recover from a single packet loss. This means that TCP cannot fully utilize the available bandwidth. The Proceedings of the 6th WSEAS International Conference on Multimedia, Internet & Video Technologies, Lisbon, Portugal, September 22-24, 2006 116 3.4 The planned future to congestion control in the current TCP. We improvements: compare TCP with FAST TCP and also discussed Token-based FAST TCP. We also Measurement of Backward Queuing check its performance in ATM Networks. delay to avoid reverse path Steven Low and Co. did a fantastic job in congestion affecting throughput. this research area. We studied their work as Detection of route change. well. Also some benefits, achievements and Tuning of Socket Buffers. some planned future improvements are also Reducing number of ACKs described. And the bandwidth-hungry processed to improve CPU entertainment industry is also looking at Fast utilization. TCP. Introduce 'TCP Friendliness' parameter to control sender References: aggressiveness in lossy  Jing WU, Peng ZHANG, Tao DU, Jian environments. MA and Shiduan CHENG, “Improving TCP SACK processing optimization . Performance in ATM Network by the Fast TCP Flow Control,” International Conference on Communication Technology 4. Conclusion ICCT’98, October 22-24: 1998 Beijing, FAST TCP is an alternative congestion China control algorithm in TCP. Lot of people  C. Jin, D. X. Wei, and S. H. Low, “TCP already worked on this subject. Their FAST: Motivation, Architecture, research work and results yield a new Algorithms, Performance,” Proc. IEEE dimension for the entertainment and INFOCOM, Mar. 2004, scientist eras. TCP flow control mechanism http://netlab.caltech.edu can only indirectly detect congestion by  Qian Wang, Jing Wu, Shiduan Cheng, keeping track of how many packets are lost, Jian Ma2, “Fast TCP Flow Control with congestion control has to be initiated after Differentiated Services” packet losses due to congestion have already  F. Peng, B. Wei and Y. Ma, “Delay happened. Therefore, if TCP control time performance analysis of token-based fast can not be speed up, TCP will cause major TCP in window limit systems”, Proc. 9th overloads and outages on long haul International Conference on Computer networks Furthermore, the maximum Communications and Networks, 2000. window size allowed in current systems is  [ 13 Peng Zhang, Jian Ma, “ Token- not large enough to catch up with based Fast-TCP ”, Invention Report, 1999 Bandwidth-Delay Products. We studied Fast  H. Wu, Jing Wu, Keping Long, Shiduan Active queue management Scalable Cheng, Jian Ma2, “TCP Enhancement: Transmission Control Protocol Token based Fast-TCP delay algorithm” comprehensively. FAST TCP uses the same  W. Qian, W. Jing, C. Shiduan and Ma packet sizes as regular TCP, the hardware Jian, “Differentiated Service Fast-TCP that carries messages around the net will still Policy for Resource Management” work. The difference is in software and  F. Peng, C.M. Leung, “Performance hardware on the sending computer, which Analysis of Token-based Fast TCP in continually measures the time it takes for Systems Supporting Large Windows” sent packets to arrive, and how long  C. Jin, D. Wei, S. H. Low, “FAST TCP: acknowledgements take to come back. This From Theory to Experiments” reveals the delays on the line, giving early http://netlab.caltech.edu/FAST/ December 6, warnings of likely packet losses. The FAST 2003 TCP software uses this to predict the highest  http://netlab.caltech.edu/FAST data rate the connection can support without  http://newscientist.com losing data. We discussed the issues related  http://pr.caltech.edu/media
Pages to are hidden for
"TCP Enhancement_ Fast Active Queue Management Scalable "Please download to view full document