Documents
Resources
Learning Center
Upload
Plans & pricing Sign in
Sign Out

a modified newreno tcp in lossy networks

VIEWS: 86 PAGES: 10

a modified newreno tcp in lossy networks

More Info
									A Modified TCP-NewReno in A LAN/WAN with A Lossy Link
Zhu Jing, Li Zhengbin, Niu Zhisheng Dept. of Electronic Engineering TSINGHUA UNIVERSITY Beijing 100084 China Email: {zhuj, Lizb}@atm.mdc.tsinghua.edu.cn niuzhs@mail.tsinghua.edu.cn

Abstract: In this paper, we propose a modified TCP-NewReno in LAN/WAN with a lossy link such as wireless network and ATM network, where packet loss occurs because of unreliable data link layer or imperfect physical transport layer. Much research has show that the scheme of TCP-NewReno or the scheme of reducing fast-retransmit threshold can effectively improve the TCP performance in lossy networks. However, through simulation, we show that these two schemes do not consider some unfavorable factors such as out of order packets delivery and retransmission losses. Through simulation and analysis, we give some changes to TCP-NewReno. Simulation results show that our change works well for both LAN connections and WAN connections Key Words: TCP/IP, wireless networks, transport protocol, Internet. 1. Introduction TCP has been designed for networks in which segment losses/corruptions are mainly due to congestion. Upon that assumption, the retransmit scheme of TCP is conservative in order to avoid unnecessary retransmission. But with Internet extending, many emerging networks access Internet, such as ATM network and the wireless network. Much search has show that there is lossy link in these networks, and packet loss occurs due to

unreliable data link layer or imperfect physical transport layer in these networks. Furthermore, in the future TCP data traffic will share the network with multimedia traffic like voice and video, which may lead to transient losses to the TCP traffic. So current TCP is not adaptive to current network environments, and optimizing protocol in transport layer is necessary and important. Much research has show that, in order to improve TCP throughput in lossy networks, it is most important to avoid successive timeout and drastic window reductions due to random losses. For this, we need an effective retransmit scheme of TCP. Most of work done before concludes that the scheme of TCP-NewReno or the scheme of reducing the fast retransmit threshold can effectively improve TCP performance in lossy networks. But, these schemes do not consider some unfavorable factors such as out of order packets delivery and retransmission losses. In [1], TCP-OldTahoe and TCP-Reno were studied, and the authors developed a basic understanding of these schemes by considering one-way traffic over a single bottleneck link with FIFO transmission, and compared their performance in lossy networks. In [2], TCP-NewReno was included, and the authors analysed improvement of TCP end-to-end throughput by reducing fast-retransmit threshold K, but they ignored out of order packets delivery. TCP-NewReno is a good solution to multiple losses in a single

window, but can not work well when retransmitted packets are lost. In [3], the analytical framework was developed to study the effect of TCP parameters, channel characters, intermittent connectivity due to handoff on TCP throughput. In this paper, we study the impact of retransmission losses. Through simulation we conclude that although probability of retransmission losses is a little once happens TCP throughput significantly degrades. Upon that, we give some changes to TCP-NewReno, and we compare our change with current TCP on throughput at high BER. This paper is organized as follows, In Section 2, we show the retransmit scheme of TCP-NewReno and analyse the failure of TCP-NewReno. In Section 3, we give our proposal. In Section 4, we show simulation results, and compare our modified TCP-NewReno with current TCP on throughput in both LAN and WAN. In Section 5, we conclude our work. 2.The Retransmit Scheme of TCP-NewReno In [3], the author has revealed that the reasons of throughput degradation of TCP over lossy networks specially, 1) False congestion control triggered by random informantion losses. 2) Multiple losses/corruptions in a single window. 3) Losses of retransmissions at high BER. The authors in [3] also show that the current fast retransmit/fast recovery algorithms inTCP-Reno do not function well when multiple losses in the same window occur. Only a few of the lost packets may be recovered by fast retransmit algorithm. The rest are often recovered by slow-start after retransmission timeout, usually with a long idle period (due to timer backoff). That results

in significant performance degradation. TCP-NewReno is a good solution to multiple losses in a single window. If some of the packets which are transmitted before the fast retransmit lose, the acknowledgement for the fast-retransmitted packet will acknowledge some but not all of the packets transmitted before the fast retransmit. That ACK packet is called a partial acknowledgment packet. TCP-NewReno will retransmit the indicated packet without delay when receiving the partial acknowledgment packet. In [2], the authors conclude that TCP-NewReno can work well when there are only multiple losses/corruptions in a single window at intermediate BER, which has been configured by our simulation. But when retransmission losses occur at high BER, the lost packets which are transmitted after the fast retransmit cannot be recovered quickly by TCP-NewReno and they also cannot be fast-retransmitted because of no enough duplicated ACKs. Therefor successive timeout is inevitable. That leads to under-utilization of network resources and long idle period. In order to explain that clearly, we give an example (see Fig. 1). We can see that successive timeout leads to a long idle time and low value of both the slow-start threshold and the congestion window size. All that will degrade TCP throughput significantly. 3. Our Modification to TCP-NewReno Reviewing of development of TCP retransmit scheme, we find that the effective way is speeding up the reaction of TCP to packet losses. But current TCP is conservative in lossy networks with high BER. In our modification to TCP-NewReno, the fast retransmit and the timeout retransmit are both called delay-retransmit. When the

delay-retransmitted packet is acknowledged, the packet sent before it should also be acknowledged if no losses happen. So, we consider those that are not acknowledged as lost, and retransmit them at once. If the delay-retransmit is false triggered when no losses happen because of underestimation of RTT, our change may lead to unnecessary retransmit so that effective throughput degrades a little. However, we consider it as the event with a little probability because of coarse estimation of timeout period that is often longer than exact timeout period in the real world. Our change to TCP-NewReno is that, before any delay-retransmit we record the highest sequence number by “H”. If acknowledgement of the delay-retransmitted packet does not cover “H”, we retransmit the indicated packet at once. We do not change other algorithms of TCP-NewReno. We use Time: T0: A=2 W=12 Sender #2 #3,#4,#5 #6,#7,…,#13

the example above to explain our modified TCP-NewReno(see Fig. 1). The notations used are as follows. A: the sequence number of the first packet in the congestion window W: congestion window size rto: timeout value t: initial timeout value SSth: slow-start threshold H: the highest sequence number transmitted before the delay-retransmit Compared with TCP-NewReno, our modified TCP-NewReno can speed up the reaction of TCP to packet losses and reduce the frequency of timeout in order to avoid drastic window reduction and long idol period, which leads to better performance at very high BER.

Receiver #2 lost #3 Duplicated ACK for #2 #4 Duplicated ACK for #2 #5 Duplicated ACK for #2

T1:A=2 W=12/2+3=9 SSth=12/2=6 rto = t ……

F-R #2 #2lost #6,#7,….,#13 Another seven Duplicated ACKs for #2 A=2 W=9+7=16 #14,#15,…,#17

T4=T1+t

T2:#14 T3:#16 …… T4: rto=2*t A=2 W=1 SSth=16/2=8 T5:

#14 #16 lost #15,#17 Another two Duplicated ACKs for #2 T-R #2 #2 ACK for #14 A=2 W=1+2=3 A=14 W=3+1=4

TCP-NewReno

Modified TCP-NewReno H=#17 T6’=T5 5 T6’:Retransmit #14 at once (14<H) #14 ACK(#16) A=16 W=5 T7’:Retransmit #16 at once(16<H) #18,#19,#20 SSth=8 A=18 W=6 T8: #21,#22,#23 #16 ACK(#18)

T6=T2+2*t

T7=T3+4*t

…… T6: T-R #14 A=14 W=1 SSth=2 rto=4*t A=16 W=1+1=2 …… T7: T-R #16 A=16 W=1 SSth=2 rto=8*t SSth=2 A=18 W=2 T8: #18,#19 T-R: Timeout-Retransmit F-R: Fast-Retransmit Rtt: Round trip time

T7’=T6’+Rtt #14 ACK(#16)

#16 ACK(#18)

( Fig 1 ) situation is considered, propagation delay is 4.Simulation Result and Discussion 4. assumed to be negligible, out of order packets Fig. 2 shows our network scenario. We delivery is also ignored and all packets from a model the wireless link as simply a lossy link data flow are assumed to follow the default and use the greedy source that generates data path. On the other hand, in a wide area when needed. TCP packets are fixed length network, we assume they may follow two with 512 bytes. The source is 155.2 Mbps. paths with different propagation delay ( τ Buffer size of intermediate system is 100 TCP 1=10 ms,τ2=10.1 ms)in order to study the packets. Transmit rate of the intermediate impact of out of order packets delivery on system is 2 Mbps. When a local networking TCP throughput. We also suppose that ACK

Rt=155.2 Mbps

Network Cloud (t,Prob) TCP source
default path: LAN t=0ms WAN t=10ms the second path: WAN t=10.1ms

Rt=2 Mbps Buffer=100 TCP packets Intermediate System
Lossy Link

Mobil Terminal

Fig 2 Network Scenario

Probability selecting the second path in WAN:"Prob" Propagation delay:"t"

packets are transmitted without loss since they are generally assigned a higher priority. Because ACK packets are relatively much smaller than data packet (40-byte ACK’s versus 512-byte packet), its propagation delay is much shorter. For simplification, we ignore its propagation delay. In simulation, our modified TCP-NewReno is called NewReno2. We first study the impact of both multiple losses and retransmissions losses on TCP throughput in LAN. Fig. 3b demonstrates TCP–NewReno can retransmit lost packet at once in order to recover from losses as soon as possible when multiple losses in a single window happen. When retransmission losses happen, TCP-NewReno will experience a long idle time, because TCP-NewReno can not retransmit lost packet at once, which is sent during the period between the Fast Retransmit and the Timeout Retransmit and successive timeout is inevitable (Fig. 3c). That will significantly degrade throughput. While our modified TCP-NewReno can retransmit lost packets at once to improve throughput (Fig. 3d). From Fig.3a, we can conclude that our modified-TCP speeds up TCP’s reaction to losses so that the performance of TCP end-to-end throughput improves greatly. Fig. 4a shows that, without out of order packet delivery reducing fast-retransmit threshold K can improve throughput. We also conclude from Fig. 4b that, compared with current TCP our modified TCP-NewReno gets the best performance in LAN with a lossy link. In an error free WAN, we mainly study the impact of out of order packet delivery on throughput. From Fig. 5 we see that, with increasing the probability of selecting the second path which leads to out of order packet deliveray, throughput degrades because of many unnecessary retransmissions. TCP-NewReno(K=1) degrades most greatly, while both our modified TCP-NewReno and

TCP-NewReno(K=3) work well. In a WAN with a lossy link, we compare our modified TCP-NewReno, TCP-NewReno(K=1), and TCP-NewReno(K=3). Fig 6 shows that, when the impact of out of order packet delivary is mainly at low BER. Modified TCP-NewReno and TCP-NewReno(K=3) work better than TCP-NewReno(K=1). With increasing BER, the impact of random losses become more and more important, TCP-NewReno(K=3) throughput degrades significantly so that TCP-NewReno(K=1) throughput exceeds it, and our modified TCP-NewReno still works well. 5. Conclusion Through simulation and comparison, we reveal that when out of order packet delivery is considered in a WAN the scheme of reducing fast retransmit threshold K may lead to throughput degradation. We also investigate the impact of retransmission losses on TCP-NewReno and conclude that TCP-NewReno cannot effectively recover from retransmission losses. Finally, we conclude that our modified TCP-NewReno works better than current TCP in WAN/LAN with a lossy link. In our future work, we will consider unreliable ACK packets transmit and variation of propagation delay. Reference: [1] T. V. Lakshman, Upamanyu Madhow “The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss” in IEEE/ACM Transaction on Networking. vol. 5. no. 3. June 1997. [2] Anurag Kumar “Comparative Performance Analysis of Versions of TCP in a Local Network with a Lossy Link” in IEEE/ACM Transaction on Networking. vol. 6. no. 4. June 1998. [3] Aldar C. F. Chan, Danny H. K. Tsang, Sanjay Gupta “Performance Analysis of TCP in the Presence of Random Losses/Errors”

IEEE Globe Com ’98, Sydney, Australia, 1998

Fig. 3a Simulation traces comparing TCP-NewReno and our modified TCP-NewReno

Fig.3b A close-up view between 1.09 to 1.107 sec of simulation traces showing how TCP-NewReno effectively recovers from multiple losses in a single window in LAN with a lossy link. (BER=1e-4)

Fig. 3c A close-up view between 1.15 to 1.25 sec of simulation traces showing how TCP-NewReno can’t effectively recover from retransmission loss in LAN with a lossy

link.(BER=1e-4)

Fig. 3d A close-up view between 0.94 to 0.99 sec of simulation traces showing how our modified TCP-NewReno effectively recovers from retransmission losses in a single window in LAN with a lossy link. (BER=1e-4)

40000 35000 30000 25000 20000 15000 10000 5000 0 1.00E06 1.00E05 1.00E04 1.00E03 Reno(K=1) Reno(K=3)

Effective Throughput

Bit Error Rate

Fig. 4a Throughput of versions of TCP versus bit error rate in LAN with a lossy link. (K is the fast-retransmit threshold)

40000 35000 30000 Reno(K=3) 25000 20000 15000 10000 5000 0 1.00E06 1.00E05 1.00E04 1.00E03 NewReno(K=3) NewReno2(K=3)

Effective Throughput

Bit Error Rate

Fig. 4b Throughput of versions of TCP versus bit error rate in LAN with a lossy link. (K is the fast-retransmit threshold)

40000 35000

Effective Thoughput

30000 25000 20000 15000 10000 5000 0 1.00E-05 1.00E-04 1.00E-03 NewReno(K=1) NewReno(K=3) NewReno2(K=3) 1.00E-02 1.00E-01

Out of Order Packets Delivery Probability

Fig. 5 Simulation results which demonstrate the effect of out of order packet delivery on TCP throughput in a error free WAN.

40000 35000 30000
Effective Throughput

NewReno2(K=3) NewReno(K=3) NewReno(K=1)

25000 20000 15000 10000 5000 0 1.00E-07 1.00E-06 1.00E-05 1.00E-04 1.00E-03 Bit Error Rate of Lower Layer Protocol

Fig. 6 Throughput of versions of TCP versus bit error rate in WAN with a lossy link. (K is the fast-retransmit threshold)


								
To top