Improving TCP throughput using forward error correction

Transcription

1 This article has been accepted and published on J-STAGE in advance of copyediting. Content is final as presented. IEICE Communications Express, Vol., 1 6 Improving TCP throughput using forward error correction Yurino Sato 1a), Hiroyuki Koga 1, and Takeshi Ikenaga 2 1 Graduate School of Environmental Engineering, The University of Kitakyushu 1-1 Hibikino, Wakamatsu, Kitakyushu, Fukuoka , Japan 2 Graduate School of Engineering, Kyushu Institute of Technology 1-1 Sensui-cho, Tobata, Kitakyushu, Fukuoka , Japan a) yuri@net.is.env.kitakyu-u.ac.jp Abstract: Packet losses significantly degrade TCP performance in high latency networks. This is because TCP keeps transmission rates low while the lost packets are recovered. To prevent this problem, retransmissions must be kept as low as possible. Therefore, we propose a scheme to apply FEC technology to the entire TCP operation in order to prevent degradation of transmission rates as well as improve throughput. Since a simple application of FEC technology might not work eftively, we consider the ways to control the return of ACKs and redundancy level. We show the eftiveness of this approach through simulation evaluations. Keywords: TCP, FEC, Redundancy Control, Congestion Control Classification: Network References [1] A. Vishwanath, V. Sivaraman, M. Thottan, and C. Dovralis, Enabling a bufferless core optical network using edge-to-edge packet-level FEC, IEEE Transactions on Communications, vol. 61, no. 2, pp , 213. DOI:1.119/TCOMM [2] Y. Sohn, J. Hwang, and S. Kang, Adaptive packet-level FEC algorithm for improving the video quality over IEEE networks, SERSC International Journal of Software Engineering and Its Applications, vol. 6, no. 3, pp , 212. [3] M. Sakakibara, Y. Tanigawa, and H. Tode, Highly reliable TCP transfer method with error correction technology, Proc. International Conference on Multimedia, Information Technology and its Applications, pp , Aug. 29. [4] Y. Huang, S. Mehrotra, and J. Li, A hybrid FEC-ARQ protocol for lowdelay lossless sequential data streaming, Proc. International Conference on Multimedia and Expo, pp , Jun./Jul. 29. DOI:1.119/ ICME [5] V. Sharma, K. Ramakrishnan, K. Kar, and S. Kalyanaraman, Complementing TCP congestion control with forward error correction, Proc. IFIP International Networking Conference, pp , May 29. DOI: 1.17/ IEICE 216 DOI: /comex.216XBL158 Received September 2, 216 Accepted October 3, 216 Publicized October 24, Earlier version of this letter was presented at IEEE ICN13 Poster Session, Oct. 1

2 [6] T. Tsugawa, N. Fujita, T. Hama, H. Shimonishi, and T. Murase, TCP- AFEC: An adaptive FEC code control for end-to-end bandwidth guarantee, Proc. International Packet Video Workshop, pp , Nov. 27. DOI:1.119/PACKET Introduction A communication over the Internet still commonly uses TCP as a reliable transmission protocol in today s global and broadband networks. However, TCP cannot utilize such broadband networks, particularly in high latency environments. This is because TCP keeps the transmission rate low while the lost packets are recovered. To prevent this problem, retransmissions must be kept as low as possible. One efficient way to prevent packet losses is a forward error correction (FEC) technology that enables a sender to transmit packets with redundant information to correct transmission errors at a receiver. To use it eftively, the amount of redundant information must be appropriately determined according to network conditions. Therefore, FEC technology is difficult to adapt to TCP communication where transmission rates change a lot because it is harder to select an appropriate redundancy level in such an environment. For this reason, although there have been few studies on the application of FEC technology to TCP, there have been studies on TCP restrictively using FEC technology. In this study, we propose a scheme to apply FEC technology to the entire TCP operation in order to prevent degradation of transmission rates as well as improve throughput. However, a simple application of FEC technology to TCP operation might not work eftively. It causes unnecessary retransmission due to the reception of duplicate ACKs even if recovery is successful. Therefore, we consider three ways to control the redundancy level according to transmission rates, to limit minimum redundancy, and to suppress the return of duplicate ACKs. We show the eftiveness of this scheme through simulation evaluations. 2 Related work As mentioned above, the recovery success rate depends on the redundancy level. By the way to control redundancy, the proposed FEC algorithms are categorized into two groups: keeping constant redundancy level independent of network conditions [1, 2] and dynamically adjusting redundancy level according to network conditions [3, 4, 5, 6]. Ref. [3] proposes a method to apply FEC to TCP only during a recovery phase. FEC-ARQ [4] proposes a method to combine FEC technology and ARQ mechanisms to improve quality of streaming services. Ref. [5] focuses on the problem that it takes a long time to recover lost packets even when short-term burst losses occur in a high latency broadband environment. The proposed method uses adaptive FEC with redundancy proportional to the number of lost packets only when 2

3 Sender Congestion window size Receiver XOR R1 R1 Suppress ACKs R1 XOR Recover lost packet P4 Send ACKs P5 P : Original packet R : Redundant packet : Segment : ACK Fig. 1. Proposed scheme it detects packet losses by Explicit Congestion Notification (ECN). TCP- AFEC [6] uses FEC technology for streaming services, a method that does not greatly change the transmission rate. It controls redundancy based on a timer and a required transmission rate from applications. These works focus on TCP restrictively using FEC technology. 3 Proposed scheme Our proposed scheme which applies FEC technology to entire TCP operation controls redundancy level according to transmission rates. The redundancy decreases as the transmission rate increases considering network efficiency. This means that redundancy is high at a low transmission rate where packet losses have a significant impact. Fig. 1 illustrates the operation of our scheme. The sender creates a redundant packet XORed with the payload field of all packets in a congestion window (cwnd) and transmits the redundant packet following the original packets. Namely, our scheme forms a group of a redundant packet and corresponding original packets. The receiver thus can use the redundant packet to recover a lost packet within the cwnd. This scheme cannot recover lost packets when two or more packet losses occur within a cwnd. In that case, the sender retransmits the lost packets through the original TCP operation. As mentioned above, FEC mechanisms cannot work eftively if simply applied to TCP operation. When the redundancy is particularly low, FEC mechanisms might cause unnecessary retransmission due to the reception of duplicate ACKs even if recovery succeeds and might not recover lost packets 3

4 eftively. Therefore, our scheme has two mechanisms to suppress duplicate ACKs and control minimum redundancy. First, our scheme suppresses returning duplicate ACKs until it is determined that a lost packet cannot be recovered; i.e., when another original packet or a redundant packet in a group is lost. Second, our scheme introduces an upper limit to group size fmax to prevent inefficient operation due to too low redundancy. 4 Simulation model To investigate the efficiency of our scheme, we evaluated it through simulation using Network Simulator ns-2 ver after implementing our scheme. In this simulation, a sender communicates with the corresponding receiver through two routers. We assumed that random packet loss of.1 1% occurs over the link between routers. The link has a bandwidth of 1 Mb/s and a delay time of 1 3 ms. Other links have a bandwidth of 1 Mb/s and a delay time of 1 ms. The sender employing TCP with selective ACK option () and the proposed scheme transmits continuous data packets to the receiver. In the proposed scheme, we compared the performance of three methods: method where the redundancy is adapted according to cwnd, +fmax n methods which sets the upper limit of group size to n packets in addition to the method, and -dup method which does not suppress duplicate ACKs. In this simulation, the fmax value varied from 1 to 5. To evaluate the eftiveness of our scheme, we focus on the average throughput, eftive recovery rate, and number of TCP timeouts and fast recoveries. The eftive recovery rate is defined as the ratio of recovered packets to redundant packets; e.g., the rate of 1% means that the redundant packets of 9% are not utilized eftively. We evaluated the average values over 4 simulation runs with different random seeds. 5 Simulation results Fig. 2 shows the average throughput performance when the delay time, packet loss rate, and fmax vary. From Fig. 2(a), our scheme achieves higher throughput than method. Especially, +fmax 1 method achieves excellent throughput even as the delay time increases although throughput deteriorates with other methods. On the other hand, -dup method attains as low throughput as method. Namely, FEC technology cannot work eftively if simply applied to TCP operation. Fig. 2(b) shows that our scheme achieves higher throughput than method in a wide range of packet loss rates. Our scheme with large fmax achieves good throughput at a low packet loss rate, while that with small fmax does it at a high packet loss rate. To investigate the eft of fmax on throughput performance in detail, the average throughput of +fmax n methods is shown in Figs. 2(c) and 2(d), when the fmax value varies. These results indicate that the packet loss rate has a large impact on the eft of fmax compared with the delay time. Since large fmax means low redundancy, it degrades throughput significantly 4

5 Delay time [ms] fmax 1 +fmax 1 +fmax 4 -dup +fmax 1 +fmax 1 +fmax 4 -dup (a) Eft of delay time (Packet loss rate =.5%) (b) Eft of packet loss rate (Delay time = 5 ms) fmax fmax 1 ms 5 ms 1 ms 2 ms 3 ms.1%.5%.1%.5%.1%.5% 1% (c) Eft of fmax (Packet loss rate =.5%) (d) Eft of fmax (Delay time = 5 ms) Fig. 2. Throughput performance Number of fast recoveries fmax 1 +fmax 1 +fmax 4 -dup Number of timeouts fmax 1 +fmax 1 +fmax 4 -dup (a) Number of fast recoveries (b) Number of timeouts Eftive recovery rate [%] fmax 1 +fmax 1 +fmax 4 -dup (c) Eftive recovery rate Fig. 3. Analysis of the characteristics at a long delay time and high packet loss rate. On the other hand, small fmax causes inefficient transmission due to high redundancy. Therefore, an appropriate fmax value is 5 1 in this environment. 5

6 Let s investigate the reason for the above improvement. The number of fast recoveries and timeouts, and eftive recovery rate of each method are shown in Fig. 3, when the delay time is set to 5 ms. From Fig. 3(a), we can see that +fmax n and methods reduce the number of fast recoveries much more than the other methods do. However, +fmax 4 and methods cause many timeouts at a high packet loss rate as shown in Fig. 3(b). This is because the suppression time of duplicate ACKs increases as the group size increases in these methods. It causes many timeouts as well as a little throughput improvement. On the other hand, +fmax 1 method drastically reduces the number of fast recoveries and timeouts, although the eftive recovery rate of it is very low as shown in Fig. 3(c). Clearly, +fmax 1 method can recover lost packets eftively and prevent retransmission by redundant packets. Consequently, our scheme improves throughput performance regardless of the delay time by setting an appropriate redundancy level according to packet loss rates. 6 Conclusion Packet losses significantly degrade TCP performance in high latency networks. To improve TCP performance in such networks, we proposed a scheme to apply FEC technology to the entire TCP operation. The proposed scheme consists of three mechanisms to control the redundancy level, to limit minimum redundancy, and to suppress duplicate ACKs. Simulation evaluations show that the proposed scheme improves throughput significantly by suppressing the return of duplicate ACKs and controlling minimum redundancy, especially in high latency environments, although FEC technology cannot work eftively when simply applied to TCP operation. In future work, we will consider a scheme to determine the appropriate redundancy level for network conditions and to more eftively recover lost packets in a real environment, such as where bust packet losses occur. Acknowledgments This work was supported in part by the Japan Society for the Promotion of Science, Grant-in-Aid for Scientific Research (B) Number 16H286. 6