Implementation of A Loss Differentiation Algorithm For Differentiation of Non-congestion Losses From Congestion Losses For Improving TCP Performance Over Wireless Networks

Implementation of A Loss Differentiation Algorithm For Differentiation of Non-congestion Losses From Congestion Losses For Improving TCP Performance Over Wireless Networks

Nikita B. Dalwadi

ME Scholar

Presented in this paper is the simulation based study of the TCP performance issues over wireless network . The paper includes description of various LDA algorithms with its strengths and weaknesses. The author has implemented LDA sche me named spik e scheme used for discriminating between wireless and congestion losses at TCP The entire discussion open up the grey area where an effort can be put up to further improve TCP performance particularly in presence of wireless losses.

.

1. Transport layer is one of the core layers in network protocol stack. Purpose of transport layer is to provide transparent and reliable transfer of data between end to end processes relieving upper layers from any concern of reliability. It ensures the reliable arrival o f messages by providing error checking mechanisms and data flow controls. It turns the unreliable and elementary service provided by lower layers into a reliable service. It ensures that whole message arrives intact and in order at the receiver. At a mo ment there can be more than one application trying to send and receive data. The transport layer should be capable enough for enabling all these applications to send and receive data using the same lo wer-layer protocol imp le mentation.

   TCP is a reliable byte stream based connection oriented transport layer protocol.Traditional Transmission Control Protocol (TCP) is designed and optimized for wired networks under the assumptions that wired media have very low loss and propagation time [6]. It assumes that every packet loss is due to congestion on network and takes congestion control

   measure by reducing window size [3], [4], [5]. But in case of wire less networks acknowledgement may be delayed due to low wire less bandwidth or wire less link can go down without any warning. TCP can take this scenario as congestion and retransmits lost packets which reduces network bandwidth utilisation and leads to poor performance. Larger window size also poses a problem of co mpetit ion between too many packets to get the same med iu m hence causing congestion and increasing end-to-end delay. There is a need to change TCP.

   TCPs performance is poorly exaggerated in wireless networks [6].This performance issues are ma inly due to inherent characteristics of both; wire less lin ks as well as TCP loss recovery[6].Wire less network lin ks are prone to, Very high BER wh ich is unpredictable, frequent path/link brea ks due to high mobility, Bandwidth constraints , Shared Broadcast Radio Channel, Resources constraints.

   Packets lost under all above conditions are unfortunately considered as a sign of network congestion. Hence TCP attempts loss recovery using retransmissions at reduced rate. It reduces its window size considering it as congestion [6]. Thus frequent losses due to all above mentioned reasons prevent TCP to transmit more number of packets as much of the time is wasted in loss recovery and throughput (number of packets sent from source to destination in specific time) of the network degrades. As a result more than 60% of the network capacity rema ins unutilized .One of the approaches preferred by the researchers is to decouple TCP congestion control fro m its loss recovery for non – congestion losses using proper Loss Differentiat ion Algorith ms. LDA are used to differentiate congestion losses from non -congestion losses.

   The paper is organized as follo ws. Section 2 describes various related work along with description of various LDA schemes section 3 shows analysis of TCP New Reno (Simu lations are carried out using Network Simu lator ns v.2.35 ) , section 4 includes

   imple mentation of spike scheme with TCP Ne w Reno and section 5 concludes the paper.

2. Many schemes are classified into three basic groups, based on their fundamental philosophy: end -to-end proposals, split-connection proposals and link-layer proposals.

   1. Link Layer Solutions

      All these protocols ensure that link layer corrected all errors over wireless interface, hence eliminating need for error handling at TCP layer. Various approaches ma ke use of either link-level automatic retransmission request or forward error correction codes or hybrid of both. Main objective of all approaches is to hide losses other than congestion from the sender TCP The main advantage of emp loying a lin k-layer protocol for loss recovery is that it fits naturally into the layered structure of network protocols. The main concern about lin k-layer protocols is the possibility of adverse effect on certain transport-layer protocols such as TCP. Fe w of proposed schemes are: Snoop TCP [13], TCP- Unaware Lin k Layer[13].

   2. Split Approach Based Solutions

      This approach involves splitting of TCP connection based on the wired or wire less domain. The trad itional TCP can be used for the wired part and some optimized version for wire less part. The intermediate agent is known as access point (AP) and acts as proxy for M N. Thus the connection is split into two distinct connections, one between MN & AP and another between the AP & Correspondent Node. However it incurs extra protocol overhead and violates end -to-end semantics of TCP. It a lso complicates handoff due to state informat ion handling at AP where the protocol is split. Few of the algorith ms proposed are: Indirect TCP [13], Mobile TCP [13].

   3. End to End Solutions

      This category of techniques assumes that the entire setup is to be changed. Idea is to ma ke sender aware that some losses are not due to congestion and, thus, avoid congestion control when not needed. Protocols under this approach make use of selective acknowledge ment for e rror recovery. To d istinguish between congestion and other losses Explic it Loss Notification is used. Advantage of this scheme is that it ma intains end to-end semantics and incurs no extra overhead at AP for protocol processing or handoff. But at the same time it requires TCP to be modified. Few of

      suggested protocols are: Explic it Loss Notification, Wireless TCP, TCP New Reno [4] [13].

      Selective Acknowledgement and TCP Ne w Reno [4] [13].

      Although a wide variety of TCP versions are used on the Internet, the current de facto standard for TCP imple mentations is TCP-Reno. We ca ll this the E2E protocol, and use it as the standard basis for performance co mparison The E2E-NEW RENO protocol improves the performance of TCP -Reno after mu ltip le packet losses in a window by re ma ining in fast recovery mode if the first new acknowledgment received after a fast retransmission is partial, i.e , is less than the value of the last byte transmitted when the fast retransmission was done. This method enables the connection to make progress at the rate of one segment per round trip time, rather than stall until a coarse timeout. Along with some researchers have a proposed end to end loss differentiation a lgorith m which diffe rentiates error losses and congestion losses. So using this we can differentiate the type of losses and then make changes at the sender transmission rate. Following are various base loss differentiat ion algorith ms.The Loss differentiat ion decision can be obtained on TCP variab le states namely congestion window (cwnd is the ma ximu m nu mber o f packets which can be sent in one transmission), slow start threshold. The Loss differentiation decision can be obtained on TCP variab le states namely congestion window(cwnd is the ma ximu m nu mber o f packets which can be sent in one transmission), slow start threshold (ssthresh) and Round Trip Time (RTT).These class of LDAs uses delay measures to estimate the congestion status. Higher RTT values are supposed to be the effect of increased queuing delay over the network. The follo wing first three sections describes the base algorithms and next three sections describes the enhanced LDAs[14] [15].

   4. Biaz scheme

      This scheme uses packet inter-arrival time to diffe rentiate between the loss types. It uses the time (Ti) between the arrivals of last in sequence packet received by the receiver before a loss happened (Pi) and the first out of order packet received after the loss (Pi+n+1) , where n is the no. of packet loss[1] , [12].

      After the arrival of Pi if Pi+n+1 arrive at the time e xpected then the losses are assumed to be wireless loss and if it a rrive much earlier or later then it is assumed to be congestion loss [1] [12].

   5. Zig Zag scheme

      Zig Zag [1] c lassifies losses as wireless based on the no. of losses n, and the difference between rtti (RTT of the ith packet) and its mean (rttimean) (mean value of RTT for ith packet). rttdev is the deviation in RTT value fro m previous.

      A loss is classified as a wire less loss if

      n=1 and rtti < rttimean – rttdev (1) OR n=2 and rtti < rttimean – rttdev /2 (2)

      OR n=3and rtti < rttmean (3)

      OR n > 3 and rtti < rttimean – rttdev /2 (4) OR e lse the loss is classified as congestion loss.

   6. Spike scheme

      The spike scheme d ifferentiates among degree of congestion and not explicit ly differentiates congestion loss from wireless loss [1] [10]. The Relat ive One way Trip Time (ROTT) is a measure of the time a packet takes to travel fro m sender to receiver. ROTT is used to find the state of current connection. If the connection is in the spike state losses are assumed to be due to congestion otherwise losses are assumed to be wireless.

      Spike state:On receipt of the packet with sequence number i, if the connection is currently not in spike state, and the ROTT for packet i e xceeds the threshold Bspikestart then the algorithm enters the spike state [1] [10]. Otherwise if the connection is currently in the spike state and the ROTT for packet i is less than the second threshold Bspikeend the algorithm leaves the spike state [1] [10].

      Bspikestart=rottmin+*(rottmax-rottmin) (5)

      Bspikeend=rottmin+*(rottmax- rottmin) (6)

      Where, = 0.5 and = 0.33

      This research work uses TCP Ne w RENO and LDA spike scheme for the imp rovement of TCP New Ren o

3. The simu lation setup contains 3 nodes N0, N1 & N2 with (N1) as a router. N0 is the TCP and UDP sender and N2 is the receiver for both. An error model is placed in between N1 and N2. The setup is shown below in fig.3.1.

   N0 N1 N2

   Fig.3.1simulat ion setup

   Case 1: Without congestion and error

   In this case simply a TCP runs an ftp application between N0 and N2 for 10 secs. In the graph cwnd is the congestion window wh ich indicates the ma ximu m number of packets which can be sent in one transmission. Throughput is the total number of packets sent in specific t ime.

   Fig 3.2a graph of cwnd for case1

   Fig 3.2b graph of throughput for case1

   Case 2: With congestion

   For the same above network there is a TCP connection between N0 and N2wh ich runs an application of ftp for 10 sec a UDP between N0 and N2 runs for the same time. UDP is a real time applicat ion which introduces congestion by flooding the link with additional packets The results for throughput and congestion window (cwnd) are as shown below:

   Fig 3.3a graph of cwnd for case2

   Fig 3.3b graph of throughput for case2

   Case 3: W ith error

   For the same above network TCP runs for 10 sec and an error model is placed between N1 and N2. The results for this are shown below. Because of the error model packets are dropped deliberately and so the window decreases.

   Fig 3.4a graph of cwnd for case3

   Fig 3.4b graph of throughput for case3

   Summary of results

   With congestion throughput of the network decreases. For the case of network in which there is only error TCP acts the same for congestion which proves that even though there is no congestion TCP decreases the congestion window and throughput decreases. Thus there is a fa lse triggering of congestion control measures of TCP which should be eliminated for increased throughput. This is shown in the next section in imp le mentation of LDA with TCP Ne w Reno.

4. Implementation o f LDA spike scheme with TCP New Reno

   The network diagram for the simu lation is shown in the

   figure 4.1. The network consists of five nodes of which b0, so, r0 are the senders of UDP, TCP and UDP1 respectively, n 1is the router and s1 is the receiver for all senders. The data rate and delay between b0, s0 and r0 to n 1 is 11mb and 20ms. And between n1 and s1 is 2mb and 100ms. Simu lation run time is 50ms. An error model is placed between n1 and s1. Simu lations are carried for three cases that is with error only, with congestion only and with both error and congestion together. For congestion UDP and UDP1 a re run for 5 to 25ms along with TCP. For the imple mentation of spike scheme RTTmin (Round Trip Time) and RTTma x a re ca lculated on arrival of each ACK fro m which Bspikestart and Bspikeend is calculated according to the equation 5 and 6 ment ioned above.

   b0

   n1 s1

   s0

   r0

   Fig, 4.1 Simulation setup

   After that on detection of packet drop if the loss is due to congestion than TCP reacts as norma l TCP reducing congestion window but if it is due to error loss than the cwnd is kept unchanged if 3 dupacks are received till the recovery of the lost packets. But if the loss is error loss and there is time out than cwnd is reduced to one. The graphs of congestion window (cwnd), readings for average cwnd and average throughput are shown in the figures below. A ll co mparison is between TCP Ne w Reno and TCP New Reno with LDA spike scheme .

   Simu lation results:

   Fig 4.2cwnd for TCP new Reno with error only

   Fig 4.3cwnd for TCP new Reno with LDA with error only

   Table 4.1 comparison of average throughput and average cwnd for TCP new Reno And TCP new Reno with LDA with error only.

   Error rate	TCP New Reno		TCP New Reno with LDA
   	Average- throughput	Average- cwnd	Average- throughput	Average- cwnd
   0	212188	46	212188	46
   0.001	161765	34.03	201781	41.92
   0.002	141022	30	201860	43.83
   0.003	110962	25.93	191119	40.23
   0.004	97689.3	20.79	200516	40.47
   0.005	97689.3	20.79	187085	39.34
   0.006	97387.8	20.82	190905	39.81
   0.007	89019.4	19.14	189457	39
   0.008	78312.1	17.18	168697	36.79
   0.009	75241.9	16.48	153408	33
   0.01	62370.4	14.27	144986	33

   Fig 4.5 graph of average throughput of TCP new Reno and TCP new Reno with LDA with differet error rates with error only

   Fig 4.5 graph of average cwnd of TCP new Reno and TCP new Reno with LDA with different error rates with error only

   Table 4.2 Comparison of average throughput and average cwnd for TCP new Reno and TCP new Reno with LDA with congestion only.

   	TCP New Reno		TCP New Reno with LDA
   	Average- throughput	Aver age- cwnd	Average- throughput	Avera ge- cwnd
   UDP	174659	39.9	174659	39.97
   UDP 1	173940	40.04	173940	40.4
   Both	149829	37.43	149829	37.43

   Table 4.3 comparison of average throughput and average cwnd for TCP new Reno and TCP new Reno with LDA with both congestion error.

   Error rate	TCP New Reno		TCP New Reno with LDA
   	Average- throughput	Average- cwnd	Average- throughput	Average- cwnd
   0.001	156019	34.55	170016	39.03
   0.002	113439	25.79	169568	38.12
   0.003	114404	26.41	169837	38.16
   0.004	110734	25.64	157032	35.93
   0.005	81715.6	17.37	162974	34.92
   0.006	73371.8	16.1	160686	33.59
   0.007	67559.3	14.91	152054	32
   0.008	62321.2	13.69	150901	31.32
   0.009	53792.7	12.01	149012	33.2
   0.01	58814.9	13.01	144798	31.18

   Fig 4.2cwnd for TCP new Reno with both congestion and error

   Fig 4.3cwnd for TCP new Reno with LDA with both congestion and error

   Fig 4.5 graph of average throughput of TCP new Reno and TCP new Reno with LDA with different error rates for both congestion and error

   Fig 4.5 graph of average cwnd of TCP new Reno and TCP

   new Reno with LDA with different error rates for both congestion and error

   The result shown above proves that the throughput of the network increases in presence of error with the use of loss differentiat ion Algorith m. Also it does not affect the normal operation of TCP for congestion. Only during the presence of error it alte rs the window. But doing so, that is in increasing the cwnd the number of dropped packets also increases. This is due to the congestion in the network and network as well as routers capacity. This is shown in the table below.

   Table 4.4 comparison of dropped packets

   TCP New Reno		TCP New Reno with LDA
   No of Dropped Packets	Average Throughput	No of Dropped Packets	Average Throughput
   8	191954	10	211816
   10	166443	12	209727
   10	143814	13	209848
   11	126190	16	209315
   14	112117	18	208337
   14	97000.4	20	208819
   16	78049.8	22	207685
   18	65829.3	24	207369
   20	60328.2	25	196258

5. The paper presents the simulation of a network for TCP new Reno and the same with including LDA spike scheme. It has been shown that the performance of TCP new Reno improves for error losses with the use of LDA. Thus it can be used for networks having various other losses in the wire less networks to eliminate fa lse triggering during non-congestion losses and improve the throughput of the network. But a lso in keeping the cwnd unchanged for error losses it increases congestion

   losses and so the no. of dropped packets. Thus care must be taken while running the simulation for a longer duration.

1. Song Cen , Pamela C. Cosman , and Geoffrey M . Voelker

   , End to end Differentiation of congestion and wireless losses. Submission to ACM / IEEE transaction on networking, volume 11, issue 5, October 2003.

2. TCP / IP Protocol suite by Behrouz A. Forouzan. Tata M cGraw Hill Publishers

3. RFC 2581 TCP congestion control April 1999.

4. RFC 2582 TCP New Reno Modification to TCPS Fast Recovery April 1999.

5. Simulation based comparison of Tahoe, Reno and SACK TCP, Kevin Fall and Sally Floyd. S

6. TCP performance issues over wireless link George Xylomenos and George c. polyzos, Athens University of Econonics and Business, Greece Perti M ahonen and M ika Saaranen, University of Oulu, centre for wireless Communications, Finland. IEEE communications M agazine 2001, Volume 39.

7. C. Paras, J.J. Garcia-Luna-Acevez, Differentiating congestion vs. random loss: a method for improving TCP performance over wireless links, Proceedings of IEEE WCNC, pp. 90-93.2003.

8. The Network Simulator ns v.2.35.

9. S. Biaz N. Vaidya, Discriminating congestion losses from wireless losses using inter-arrival times at the receiver, in Proc. 1999 IEEE symposium on Application-Specific Systems and Software Engr. and Techn., pp. 10-17, Richardson, TX, M arch 1999.

10. Y. Tobe, Y Tamura, A. Molano. S. Ghosh and H.

    Tokuda, Achieving moderate fairness for UDP flow by path status classification, in Proc. 25 th Annual IEEE conf. on Local Computer Networks (LCN 2000), pp.252-61, Tampa, FL, Nov 2000.

11. D. Burman and I. M atta, Effectiveness of Loss

    Labelling in Improving TCP Performance in Wired/Wireless networks, Boston University Technical Report, 2002.

12. S. Biaz and N. H. Vaidya, Distinguishing Congestion Losses from wireless Transmission Losses: a Negative Result, Proc. of IEEE 7th Int. conf. on computer communications and Networks, New Orleans, LA, USA, October 1998.

13. Balakrishnan, H., Padmanabhan, V., Seshan, S. and Katz, R. A Comparison of M echanisms for Improving TCP Performance over Wireless Links, IEEE/ACM Transactions on Networking, pp. 756-769, December, 1997.

International Journal of Engineering Research & Technology (IJERT)

ISSN: 2278-0181

Vol. 1 Issue 3, May – 2012