ENSC 835: HIGH-PERFORMANCE NETWORKS CMPT 885: SPECIAL TOPICS: HIGH-PERFORMANCE NETWORKS FINAL PROJECT PRESENTATIONS Spring 2006 PROJECT TCP Fairness Analysis of CUBIC TCP Simulated by NS-2 Qing Chen E-mail qingc@sfu.ca 1
Road Map � Introduction � Advanced TCP for High speed network � Approaches � BIC and CUBIC TCP � Brief review of TCP NEW RENO � TCP Fairness � Fairness � Effects of Queuing Management � Simulation � Conclusions � Future Work � Reference 2
Introduction � High Speed Network � The bandwidth of network rises up to 10Gbps � The network covers the long distance � ESNet, Abilene � Current TCP faces difficulties in high speed network � Efficiency degrades when bandwidth-delay product increases � Oscillation problem � Problems with short flows 3
Approaches � Improve congestion control based on current TCP � HSTCP, STCP HSTCP-S. Floyd, AIMD-based, STCP-Tom Kelly, MIMD-based � Fast TCP Steven H. Low’s team, California Institute of Technology, delay based � BIC TCP, CUBIC TCP Injong Rhee and Lisong Xu, North Carolina State University � HTCP D. Leith and R.N. Shorten, Hamilton Institute. Two modes: High speed and slow speed. � Others � SABUL � XCP 4
BIC TCP Binary search � 1. When loss happens, calculate CWNDmin after loss using multiplicative decrease, CWND=CWNDmin 2. Calculate a mid point between CWNDmax (CWND just before loss) and CWNDmin 3. Set the mid point as target, Target=(CWNDmas+CWNDmin)/2 4. If Target-CWND>Max setting, CWND=CWND+Max setting 5. If Target-CWND<=Max Setting, CWND=Target, CWNDmin =Target, repeat step 2 to 5 6. If CWNDmax-CWNDmin<Min Setting, CWND=CWNDmax and Binary search completes Two Stages � Max. Probe stage � Binary search stage � Window Growth Pattern � (http://www.csc.ncsu.edu/faculty/rhe e/export/bitcp/index.htm) 5
CUBIC TCP � Cubic TCP is developed on BIC TCP � Main reasons � The window control algorithm of BIC is too complicated. � BIC TCP could be too aggressive in slow network with short round trip time. � Improvements � Use a cubic function to search CWnd � Involve a elapsed time since last loss when calculate CWnd 6
CUBIC TCP (Continue) � Algorithm � When receive an ACK ( ) ← − + = β 3 cwnd C t K cwnd K cwnd / C 3 max max � When loss happens ← β × cwnd cwnd max � Window Growth Pattern (http://www.csc.ncsu.edu/faculty/rhe e/export/bitcp/index.htm) 7
Brief Review of TCP New Reno � TCP New Reno � Traditional congestion control algorithm � AIMD � Four Stages � Slow Start � Congestion Avoidance � Fast retransmit � Fast recovery 8
TCP Fairness � TCP Fairness � When n flows connect to one link, each flow will share 1/n of total bandwidth � Fairness ratio of two flows = FR Thru / Thru ij i j � Factors that affect Fairness � RTT: Throughput ratio of two flows is inversely proportional to the ratio of their RTTs � Queuing management � Link Capability 9
Queuing Management’ Effect on TCP Fairness � RED (Random Early Detection) � Algorithm: � Queue model � Effects on fairness � Flow with larger sending rate will have higher drop possibility 10
(Continue) � Drop Tail � Algorithm � No packet drop if queue is not full � Drop all packet if queue is full � Queue model � Effects on fairness � Flow with higher sending rate will have more packets in queue 11
Simulation � Tool: NS-2.26 with CYGWIN � Topology and assumption: R1 S1 N1 N2 � Two flows have same RRT S2 R2 � What will be analyzed � Congestion window, throughput, fairness and link utility � Scenarios and cases � 1. Validation � 2. RED and Drop tail � 3. Three cases with different start time. � 4. Two different bandwidths: 20MBPS and 1GBPS 12
Validation Congestion Window Growth w.r.t. time: CWND of CUBIC CWND of New Reno 250 250 200 200 CWND (Packets) CWND (Packets) 150 150 100 100 50 50 0 0 1 446 891 1336 1781 2226 2671 3116 3561 1 369 737 1105 1473 1841 2209 2577 2945 3313 3681 Time (100ms) Time (100ms) 13
(Continue) Link Utility: The average link utility of Cubic is 99.4%, the average of link utility of New Reno is 89.4% Throughput of New Reno Throughput of CUBIC Thousands Thousands 25 25 20 20 15 15 Throughput (KB) Throughput (KB) 10 10 5 5 0 0 1 37 73 109 145 181 217 253 289 325 361 397 1 36 71 106 141 176 211 246 281 316 351 386 Time (s) Time (s) 14
Simulation Result Fairness Ratio (New Reno/CUBIC) DropTail 0.4 RED 0.35 Fairness Ratio 0.3 � Fairness: 0.25 0.2 0.15 � 1GBPS bandwidth, 0.1 0.05 0 100ms RTT 200 250 300 350 400 450 Time (s) � With RED: 0.09 Fairness Rate (Cubic to New Reno) � With DropTail: 0.14 DropTail RED 2 � 20MB bandwidth, 1.5 Fairness 100ms RTT 1 0.5 � With RED: 0.69 0 � With DropTail: 0.40 200 250 300 350 400 450 Time (s) 15
Conclusion � The performance of CUBIC is better than that of TCP New Reno in oscillation and link utility � In middle speed network, TCP fairness of CUBIC is better than that in high speed network � Fairness will be a problem for CUBIC in high speed network � Fairness of CUBIC with RED is a little bit better than that with DropTail in middle speed network but a bit worse in high speed network 16
Future Work � Analyze fairness performance with other configuration of TCP parameters and queue type � Compare fairness of other advanced TCP protocol 17
References [1] H-TCP: A framework for congestion control in high-speed and long- � distance networks, D.J. Leith, R.N. Shorten, Y.Lee, HI Technical Report, August 2005, http://www.hamilton.ie/net/htcp/ [2] Binary Increase Congestion Control for Fast, Long Distance Networks, � Lisong Xu, Khaled Harfoush, and Injong Rhee, ”, In Proceedings of the IEEE INFOCOM, March 2004 [3] CUBIC: A New TCP-Friendly High-Speed TCP Variant, Injong Rhee, and � Lisong Xu, PFLDnet 2005, February 2005 [4] H-TCP: TCP for high-speed and long-distance networks, D. Leith, R. � Shorten, Second International Workshop on Protocols for Fast Long- Distance Networks, February 16-17, 2004, Argonne, Illinois USA [5] A Study of TCP Fairness in High-Speed Networks, Junsoo Lee, Stephan � Bohacek, Joao P. Hespanha, Katia Obraczka, submitted to ICNP, 2005 [6] Congestion Control for High Bandwidth-Delay Product Network, Dina � Katabi, Mark Handley, Charlie Rohrs, ACM SIGCOMM Computer Communication Review, Volume 32, Issue 4, October 2002 18
Reference (Continue) [7] Fast TCP Motivation, Architecture, Algorithms, Performance, Chen Jin, � David X. Wei, Steven H. Law, In Proceedings of IEEE INFOCOM, March 2004 http://netlab.caltech.edu [8] BIC TCP, http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/ � [9] The Network Simulator - ns-2, http://www.isi.edu/nsnam/ns/ � 19
Recommend
More recommend
