Enhancing TCP Throughput over Lossy Links Using ECN-Capable Capable RED Gateways Haowei Bai Honeywell Aerospace Mohammed Atiquzzaman School of Computer Science University of Oklahoma 1 Outline Introduction and problem definition Proposed model for the maximum threshold of RED buffers Model verification Applying our model: Diff-C-TCP Conclusion 2 Page 1
Current TCP Congestion Control The original TCP/IP was designed for wireline networks. A TCP sender uses packet loss as an implicit signal for network congestion with the assumption that packet losses are mainly caused by congestion. Packet losses are indicated by Retransmission timer timing out The receipt of three duplicated ACKs Time out Slow-Start (W<=2xW per RTT) 3rd DUP ACK (Fast Retransmit) ssthresh reached Congestion Avoidance (W<=W+1 per RTT) 3 Fast Recovery (W<=W/2) 3rd DUP ACK (Fast Retransmit) Main Characteristics of Satellite Links High error rate (Packet losses often due to corruption) Long delay Relatively low bandwidth capacity 4 Page 2
Problems with Current TCP over Satellite Links The current TCP assumes that packet losses are mainly caused by congestion. However, in the satellite link, packet losses are predominantly coming from link errors. When packet losses are caused by link errors, TCP does not need to decrease its congestion window size. Using current TCP, the packet losses caused by corruption will be treated as the indication of network congestion, and hence TCP sender must decrease the speed, which degrades the performance. 5 Improving TCP Performance: ECN Explicit Congestion Notification (ECN) is an extension proposed to RED (Random Early Detection) which marks a packet instead of dropping it when the average queue size is between minth and maxth. q maxth minth RED q<minth: Accept minth<q<maxth: Randomly drop new incoming packets q>maxth: Drop all incoming packets ECN q<minth: Accept minth<q<maxth: Mark incoming packets q>maxth: Drop all incoming packets 6 Page 3
Advantages and Open Issues about ECN ECN mechanism is a congestion avoidance mechanism. ECN mechanism marks the packets before the congestion actually occurs. ECN when used with RED buffers reduces packet losses due to network congestion, and delays of TCP applications. Open issue: dimensioning buffer size and RED parameters. 7 Our Objective Our Problem Losses due to congestion Losses due to link corruption Our proposed solutions Zero congestion losses by optimally set RED buffer parameters Losses due to congestion Losses due to link corruption The objective of the paper Modeling maximum threshold of RED buffers to achieve zero congestion losses Diff-C-TCP: TCP behaves differently at two different types of loss 8 Page 4
Modeling max th of RED Buffer The random packet marking in flow i is described by a Poisson process with time varying rate λ i (t) = p(t)w i (t)/r i (t). p(t): Marking probability at the RED buffer at time t; Wi(t): Window size of the i-th TCP flow at time t; r i (t): round trip time of the i-th flow at time t. Accordingly, the waiting time (T i [n]) for the n-th marking event of flow i, which is given by T i [n] = Σ n k=1 X i (k), is a Gamma distributed random variable. X i (k) is the time interval between the (k - 1) and the k-th marking events for flow i. Specifically, the expected value of the waiting time for the first marking event is E[T i [1]] = 1/λ i (t). 9 Suggested Value of max th Assumptions: All TCP sources start sending at the same time; All packets are of the same size; Packet drops due to buffer overflow (instantaneous queue size is greater than the physical buffer size) is out of the scope of this paper. Suggested value of max th for achieving zero congestion loss: Q m max = minth + ( βi + ω) i= 1 where β i is the number of increases of the window size during time T i [1]; ω is the weighting factor of RED buffer. (Read our paper for more calculation details.) 10 Page 5
Validity of the suggested max th Three cases were simulated using NS2; simulation results are compared with analytical results. Good match between the analytical and the simulation. 11 Optimality of the suggested max th Three cases were put to the simulation: max th is three times of min th, as recommended by [1]; max th calculated by our model; 1 packet less than our suggested value. The value obtained from our model is the optimal value to achieve zero congestion loss. Our suggested value results in smaller buffer size and consequent queueing delay. [1] S. Floyd and V. Jacobson, Random early detection gateways for congestion avoidance, IEEE/ACM Transaction on Networking, vol. 1, pp. 397 413, August 1993. 12 Page 6
Applying Our Model With the negligible error, all losses can be attributed to random losses due to lossy links. Using our model, all network congestion losses can be eliminated (or the congestion losses are a small fraction of random losses). In lossy satellite links, losses due to link errors are more significant than losses due to buffer overflows. Diff-C-TCP (Differentiation Capable TCP) considers packet losses are mainly caused by corruption, rather than network congestion in a satellite link. In this context, the network congestion indicator is ECN, rather than timer timing out or the receipt of three DUPACKs, which are used in current TCP. 13 Diff-C-TCP Diff-C-TCP algorithm: assume all losses are caused by link corruption unless ECN_ECHO packets are found. Time out Slow-Start (W<=2xW per RTT) 3rd DUP ACK (Fast Retransmit) ssthresh reached Congestion Avoidance (W<=W+1 per RTT) Fast Recovery (W<=W/2) 3rd DUP ACK (Fast Retransmit) TCP-Reno Diff-C-TCP 14 Page 7
Simulating Diff-C-TCP Simulation tool: NS2. Satellite link config: 64 Kbps, 280 ms prop delay. Uniform random error generated on satellite link. BER ranging from 1e -7 to 1.2e -4. Packet size: 1000 bytes when BER is below 5e -5 ; 512 bytes when BER exceeds 5e -5. Ftp was used to transfer data. Satellite Workstation Workstation Workstation Workstation Ethernet Router Satellite dish Satellite dish Router Ethernet Workstation Workstation 15 Diff-C-TCP Improves TCP Throughput over Satellite Links Observations: At BER of 5e -5, the goodput of Diff-C-TCP is almost 5 times higher than that of the TCP. When the value of BER increases from 1e -5 to 5e -5, the current TCP s goodput decreases by 77%, while Diff-C-TCP s only decreases by 12%. 16 Page 8
Conclusions We developed an analytical model to estimate the optimal value of maximum threshold of the RED buffer. This value has been shown to be able to optimally achieve zero congestion loss. By eliminating (or almost eliminating) congestion losses, we proposed Diff-C-TCP to improve TCP throughput over satellite links. Simulation results have shown that Diff-C-TCP provides up to 5 times higher throughput performance than the current TCP algorithm at a specific error rate. Simulation results have also shown that Diff-C-TCP suffers much less than current TCP when the error rate is increasing. 17 Thank you for your attention!! For author information, please contact: Haowei Bai at haowei.bai@honeywell.com Mohammed Atiquzzaman at atiq@ou.edu 18 Page 9