Towards Performance Evaluation of Rate Control Protocol in Satellite Networks

Transcription

1 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 12 No: 2 37 Towards Performance Evaluation of Rate Control Protocol in Satellite Networks Y. D. Sun, Z. Z. Ji, H. Wang Abstract The research of congestion control has recently been focused on the explicit and rate-based mechanism. RCP (Rate Control Protocol) is an explicit and rate-based mechanism which emulates PS (Processor Sharing) to reduce the AFCT (Average Flow Completion Time). There are few studies considering RCP in satellite networks although RCP has some promising virtues. Therefore, this paper explored the performance of RCP thoroughly through simulations in NS2. The results of simulations show that RCP has high throughput and is not affected by BER (Bit Error Rates). But some drawbacks such as flash crowd, long response time and TCP-unfriendliness were found. Index Terms Congestion Control, RCP, Satellite Networks, Flash Crowd. S I. INTRODUCTION atellite networks can provide access to global information so as to transmit information effectively. Satellite networks can also provide various kinds of services such as the Internet, broadband multimedia communication and so on. Compared to terrestrial networks, satellite networks have a few advantages of scalability, wide coverage, multicast service and emergency communication after a catastrophe [1]. Therefore, the satellite networks will play a vital role in the future infrastructure of global information. Congestion control in satellite networks has become a hotspot as TCP/IP protocols are adopted in satellite networks. The first particular feature of the satellite network is long delay [2]. The GEO (Geostationary Earth Orbit) satellites are located 36Km high above the equator. So the signal transmitted from the satellite to the receiver on the earth has to endure 24~28ms delay, and the RTT (Round Trip Time) between the sender and the receiver is at least 48~56ms. Moreover, the transmission time in the other links, queuing delay in the routers and processing time at the nodes must be taken into account. So, the overall delay is much longer. The second is high BER which is caused by signal attenuation. The wireless signal is adversely affected by transmission distance, atmosphere and bad weather. The typical BER referred to [2] is 1-7, sometimes1-4[3] or even 1-2[4]. The Manuscript received March 15, 212. Y. D. Sun is with the School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China 151 ( sunyandong@hit.edu.cn). Z. Z. Ji, is with the School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China 151 ( jizhenzhou@hit.edu.cn). H. Wang is with the School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China 151 ( wanghui_cs@hit.edu.cn). high BER has negative impact on the performance of satellite networks. The third is bandwidth asymmetry which is common in satellite networks. The host of satellite networks sends data through one channel and receives ACKs through the other. The two channels have different bandwidth, BER and dely. For example, in GEO satellite networks, the forward satellite link which needs large transmitting power to transmit data has larger bandwidth, higher BER and longer delay, and the return link to transmit feedback information may be telephone line with lower bandwidth. The continuous ACKs are important for the feedback-driven protocol TCP to send new data. But the ACKs will be delayed or even lost when the number of ACKs in return link increases, which will negatively affect the forward link transmission. These features make the congestion control mechanisms in wired networks ineffective in the satellite networks, so many researchers make efforts to improve the current mechanisms or propose new ones. The newly proposed RCP [5] is a rate-based congestion control mechanism and approximates PS to reduce AFCT. The RCP has some characteristics as follows [6] : The RCP routers compute the flow rate based on the limit information as queue length and the integrated incoming traffic rate. The router allocates a single rate to every flow passing through it. The router doesn t need per-flow and per-packet computation. Intuitively, rate-based congestion control mechanism is different from the window-based one which must experience a series of window adjustment, so it is timesaving. Furthermore, the rate of rate-based mechanism is computed by router according to the condition of the networks, so the rate-based mechanism doesn t need to discriminate the packet loss as the window-based mechanism does. There have been few studies about the performance evaluation in satellite networks. Therefore, the paper evaluates and analyzes the performance of RCP in satellite networks thoroughly and the results can be referred to as the future work. II. RCP ALGORITHM The idea of RCP is to emulate PS to reduce the AFCT. The RCP is a rate-based and Max-Min congestion control algorithm. The RCP algorithm can be depicted as follows [5] : Every router maintains a single rate R( t ), allocate the rate to the flows passing through it and update the rate per RTT. Every packet header carries a rate field which keeps a rate. The rate field will be updated when the rate

2 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 12 No: 2 38 computed by the router is smaller than the original, otherwise unchanged. When the packet arrives to the receiver, the receiver copies the final rate into ACK to notify the sender. Moreover, the packet header also includes a RTT field to update the average of RTT. The sender sends data at the rate notified, which is the smallest along the link. The router periodically updates the rate as the equation(1) q( t) [ α( C y( t)) β ] d R( t) = R( t d) + (1) Nˆ ( t ) Where d is the average or all flows RTT, R( t d) is the latest updated rate, C is the link capacity, y( t) is the input traffic, q( t ) is the instantaneous queue length, N ˆ ( t ) is the estimate active flows at time t. The estimate is computed by the router and the router can t acquire the exact number of ongoing flows because some flows complete while the new ones join. And α and β are the parameters for stability and performance. Equation (1) indicates that the router allocates the spare bandwidth to all ongoing flows averagely if C y( t) >, and the rate of the active flows is reduced if C y( t) <. The router needs extra bandwidth to drain the queue within a RTT. The expression q( t) α( C y( t)) β represents the integrated d change in traffic, which is equally allocated to the ongoing flows. The number of ongoing flows is estimated as. ˆ C N ( t ) = R ( t d ) III. RELATED WORK In [6], the authors analyzed the accuracy of the estimated N( t ), and the experimental results indicated that the algorithm could estimate N( t) approximately no matter Internet flows or long flows. The lost packets are retransmitted in RCP just as in TCP. The packet loss is rare for RCP in wired networks because RCP can drain the queue in time. But in satellite networks, the packet loss caused by BER is common, so it is necessary to explore the performance of RCP in the environment. [5] indicated that RCP were stable if α <1. Parameter β is the trade-off between the acceptable queuing delay and the instantaneous rate. Larger β denotes smaller queuing delay, lower instantaneous fair rate, and vice versa. The impact of α and β on the system stability was explored in [7]. In [8], the buffer requirement of RCP in one single congestion link was evaluated and analyzed through modeling the arrival and departure of flows. The modeling procedure was: firstly, considering the impact of packet loss probability on AFCT; secondly, considering the relation between packet loss probability and buffer overflow probability; at last, deducing the relation between buffer overflow and buffer overflow probability respectively in large flow case and in small flow case. The conclusion that buffer sizes of 1% BDP (Bandwidth Delay Product) were sufficient for RCP to acquire good performance was drawn. It is advantageous to deploy RCP in high speed networks, but not in satellite networks. The packet loss in satellite networks derives from not only buffer overflow but BER, so the above modeling process doesn t apply to satellite networks. Flash crowd effect of RCP was explored in [9], especially when the number of flows increases sharply in a very short period of time. The paper also analyzed the response of RCP to different growth rates of flows and indicated that RCP queue was stable only if the growth rate didn t exceed a threshold. In [1], the author proved the fluid model of RCP was globally stable in the case that there was no delay. The problem of system stability and the conditions of system stability under the circumstances of delay were also discussed. Moreover, the RCP algorithm was further explored in one link and single delay model environment. In [11], the author proposed a variant of TCP which acquire α fairness in case of small buffer. The condition of local stability was also explored in the paper and the conclusion that feedback depending on the queue length could make the control of queue inaccurate. At last, the paper studied the span of required rate of RCP variant when new flows join through estimating the total number of flows but not one single. In [12], the author indicated that the requirement of RCP for bandwidth and buffer increased rapidly in some topologies, which resulted in system s instability. QI-RCP and PIQI-RCP mechanisms were proposed in the paper to solve the problem. The QI-RCP mechanism can make the system stable through adjusting the queue of the router. Based on QI-RCP, PIQI-RCP adds a PI controller in the router and a EWMA-like controller at the host. Simulations showed that QI-RCP and PIQI-RCP could acquire better dynamics and stability no matter in a single bottleneck link or multiple bottleneck links. Fig. 1. Topology of simulations IV. PERFORMANCE EVALUATION AND ANALYSIS We analyze and evaluate the performance of RCP, such as throughput, utilization, friendliness and so on in satellite networks in this section. We perform the simulations in NS2 [13] and the topology is referred to Fig. 1, where the satellite link is a typical bent pipe satellite and the default bandwidth, RTT, α, β and simulation time are 1 Gb/s, RTT, 55ms,.4,.9 and 5ms respectively. A. Performance of TCP In the satellite networks, TCP protocol is used by hosts to send data. The throughput of bottleneck link is considered in case of different BER. The results are as shown in Fig. 2. The figure illustrates that the peak throughput of TCP is only 3Mb/s when the BER is, and the convergence time is

3 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 12 No: 2 39 long. The throughput decreases sharply with the increase of BER, and the throughput is almost especially when the BER is.1. The figure also illustrates that the throughput goes up first then down. The reason is that TCP experiences slow start at the beginning, and the rate increases gradually. The sender reduces the window when packet loss occurs. Constant packet loss may result in the time out of the timer and the TCP restarts the procedures as slow start, congestion avoidance and so on. So, we can draw the conclusion that the performance of TCP in satellite networks is poor. the increase of BER because RCP makes an aggressive estimation when the packet loss increases. Fig. 4. Queue of bottleneck from to5 seconds. Fig. 2. Performance of TCP in satellite networks. Fig. 5. Throughput of RCP with different link BERs. Fig. 3. Performance of RCP with bandwidth of 1Gb/s. B. Performance of RCP In experiments, the congestion control protocol adopted by the hosts is RCP. We evaluated that the performance of RCP with different link bandwidth, BER, number of flows and parameters α and β. Impact of bandwidth The experimental result (the bandwidth is 1Gb/s) is illustrated as Fig. 3. The curves obtained with different bandwidth are exactly the same in shape, so we list one curve here. The reason for decrease of throughput at the beginning is that RCP doesn t estimate the bandwidth accurately and the queue builds up rapidly (as depicted in Fig. 4), which results in packet loss. The senders have to retransmit the lost packet. RCP acquires the more accurate bandwidth and the number of flows, the queue becomes low and the throughput approximates the link s capacity. Impact of BER The result is illustrated in Fig. 5 which indicates that the impact is limited although the BER has an impact on the throughput of RCP. The throughput increases instead with Fig. 6. Impact of the number of flows. Impact of the number of flows The result is referred to Fig. 6. and Fig. 7. We can conclude that the number of flows just has an effect on the throughput at the beginning and the throughput converges at last. The larger the number of flows is, the longer the convergence time lasts. It is because RCP does not exactly estimate the number of flows, which results in the burst of queue. The queue builds up as the number of flows increases and the necessary time to drain the queue increase. The estimate of the number of flows approximates the real value as the data is constantly transmitted, and the throughput converges finally. Impact of parameter α and β

4 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 12 No: 2 4 Parameter α and β are essential to the stability of the system. We evaluated the performance of RCP with various α and β respectively. The results are illustrated as Fig. 8, Fig. 9 and Fig. 1. Parameter β is used to adjust the rate to drain the buffer. The throughput decreases slightly when parameter β approaches 1. But, he figure indicates that the rate to drain the queue is high, packet loss is little and the delay is small especially in the case queue builds up rapidly at the start. C. Friendliness The results are illustrated in Fig. 11~Fig. 13. The throughput ratio approximates 2 when RCP flows coexist with TCP flows and the BER is. The throughput of TCP decreases sharply when the BER is not, while the throughput of RCP increases rapidly because the RCP flows seize the bandwidth. The reason is that the congestion control mechanism in TCP doesn t discriminate the packet loss and contributes the loss to congestion. So, congestion control is taken. However, RCP is a rate-based mechanism and just retransmits the lost packet when packet loss occurs. So, the rate doesn t decrease and the AFCT is just influenced slightly. Fig. 7. Impact of the number of flows. Fig. 1. Impact of parameter β (parameter α is.4). Fig. 8. Impact of parameter α (parameter β is.9). Fig. 11. Friendliness of RCP when BER is. Fig. 9. Impact of parameter α (parameter β is.9) Parameter α is the allocated ratio of the difference between input traffic and the link s capacity and determines the stability and the response time of the system. The figure indicates that the system exhibits flash crowd at the beginning when the parameter α approaches 1. The queue builds up rapidly and the system oscillates when the flows join and complete. Meanwhile, the response time increases accordingly. When parameter α is small, the stability of the system is enhanced and the response time decreases. Therefore, the parameter is chosen as the value between.2~.4 in practice. V. CONCLUSION AND FUTURE WORK A series of experiments were conducted in this paper, and we can conclude that the performance of RCP is better than that of TCP. RCP decouples the rate and the packet loss, so the BER doesn t affect the RCP negatively. But RCP has some drawbacks in the satellite networks. For example, the queue builds up sharply influence by burst traffic at the beginning, which results in packet loss. The scalability is undesired because the performance and stability of RCP mostly depends on the parameters α and β. The convergence time of RCP is almost several dozen seconds because of long

5 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 12 No: 2 41 RTT in satellite networks, so the short flows like Internet traffic can t fully utilize the bandwidth. RCP exhibits unfriendliness to TCP when RCP flows coexists with TCP flows because RCP is rate-based and TCP is window-based. To overcome the drawbacks of RCP in satellite networks, the future work will be focused on the scalability by realizing adaptively tuning the parameters of RCP, explore a new mechanism to solve the problem of flash crowd in case of burst traffic, and design a RCP-based friendly congestion control mechanism. departures, ACM SIGCOMM Computer Communication Review. vol. 39, no. 1, pp. 5-15, Dec 28. [9] F. Abrantes, J.T. Araujo, M. Ricardo, Flash crowd effect in RCP, Workshop Protocols for FAST Long-Distance Networks (PFLDnet). 28. [1] T. Voice, G. Raina, Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime, IEEE Transactions on Automatic Control, vol. 54, no.8, , 29. [11] F. Kelly, G. Raina, T. Voice, Stability and fairness of explicit congestion control with small buffers, ACM SIGCOMM Computer Communication Review. vol. 38, no.3, pp , Jul 28. [12] Saurabh Jain, Dmitri Loguinov, PIQI-RCP: design and analysis of rate-based explicit congestion control, 27 Fifteenth IEEE International Workshop on Quality of Service. pp. 1-2, Jun 27. [13] Network Simulator. Available: Y. D. Sun was born in He received B.E. and M.E. degrees in computer science and technology from Harbin Institute of Technology, China, in 24 and 26, respectively. He is currently a Ph.D. candidate in computer architecture of Harbin Institute of Technology. His research interests include computer networks architecture, wireless networks protocol design especially congestion control mechanisms in satellite networks. Fig. 12. Friendliness when BER is.1. Fig. 13. Friendliness of RCP when BER is.1. REFERENCES [1] S. Kota, R. Jain, and R. Goyal, Broadband satellite network performance, IEEE Commun. Mag. vol. 37, pp , July [2] M. Allman, D. Glover, L. Sanchez, Enhancing TCP over Satellite Channels using Standard Mechanism, RFC2488, IETF, Jan [3] R. T. Henderson, R. H. Katz, Transport Protocols for Internet-Compatible Satellite Networks, IEEE Journal on Selected Areas in Communications. vol. 17, no. 2, February [4] J. S. Stadler, J. Gelman, Performance enhancement for TCP/IP on a satellite channel, MILCOM Boston, vol. 1, pp , Oct [5] N. Dukkipati, M. Kobayashi, R. Z. Shen, et. al, Processor sharing flows in the Internet, IWQoS 25. Germany, pp , June 25. [6] N. Dukkipati, N. McKeown, Why flow-completion time is the right metric for congestion control, ACM SIGCOMM Computer Communication Review. vol. 36(1), pp , Jan 26. [7] H. Balakrishnan, N. Dukkipati, N. McKeown, et al. Stability analysis of explicit congestion control protocols, IEEE Communication Letters. vol. 11, no.1, Oct 27. [8] A. Lakshmikantha, R. Srikant, N. Dukkipati, et. al, Buffer sizing results for RCP congestion control under connection arrivals and Z. Z. Ji was born in He received Ph.D. degree in computer architecture from Harbin Institute of Technology, China, in 2. Since 23, he is a professor in computer science and technology with Harbin Institute of Technology. His research interests include advance computer architecture, parallel processing technology and computer networks. He is the director of T&R Division of Computer Hardware Fundamental School of Computer Science & Technology of Harbin Institute of Technology, and also the vice-director of Professional Committee of Computer Architecture of China Computer Federation. He anticipated in or presided many researches, such as projects of National Natural Science Foundation of China, prearranged projects of General Armament Department and projects of 863. He was awarded one Third Prize of the Science & Technology Progress by the Spaceflight Department of China one Third Prize of the Science & Technology Progress by the Science & Technology Committee of Heilongjiang Province and one 3rd-class prize of the Science & Technology Progress by Commission of Science Technology & Industry for National Defense. H. Wang was born in He received B.E. degree in computer science and technology from Hebei University of Technology in 25 and M.E. degree in computer science and technology from Harbin Institute of Technology, China, in 28, respectively. He is currently a Ph.D. candidate in computer architecture. His research interests include network modeling and congestion control in satellite networks.