TCP-Friendliness of SCTP and Concurrent Multipath Transfer (CMT) ILKNUR AYDIN ICCRG meeting @PFLDNET Nov 28-29 2010, Lancaster, PA, USA
Outline ● Background on TCP-Friendliness (TCP-F) ● TCP-F of single-homed SCTP Motivation, SCTP vs. TCP mechanics - - Experimental Framework Results and Analysis - Conclusions - ● TCP-F of SCTP-based CMT Motivation - Experimental Framework - Results and Analysis - Discussion and Conclusions -
Background: TCP-Friendliness (TCP-F) 3
Background: TCP-Friendliness (TCP-F) 4
TCP-Friendliness (TCP-F) ● “definition”: a non-TCP flow should not consume more resources than a confirming TCP flow under the same conditions + implement some form of congestion control mechanism Bandwidth consumed by a TCP flow* = 1.22 x MSS / RTT x sqrt(loss) * by Mahdavi and Floyd (1997), revised by Padhye (1998) and others later on 5
SCTP QualNet Module ● Comprehensive SCTP simulation module for the QualNet simulator 6
TCP-Friendliness of single-homed SCTP
SCTP vs. TCP Mechanics ● Transport Protocol Overheads - Transport PDU headers - Message-based (SCTP) vs. byte-based transmission (TCP) - Transport ACKs ● Congestion Control Mechanisms - SCTP is “similar” to TCP but already has some of the TCP enhancements (SACK, ABC, initial cwnd size, … ) Hypothesis: SCTP throughput may be better than TCP’s under the same conditions. 8
Experimental Framework 5, 10, or 20 Mbps 45 msec ● Flow: a greedy application over SCTP or TCP ( flow 1 from S1 to D1, flow 2 from S2 to D2 ● Case-I : Two flows start at the same time (how two flows grow together?) ● Case-II : Latter flow starts after the earlier is at steady-state (how one flow gives way to another flow?) ● Metrics : Throughput, Transport Load, Goodput, Fairness Index, Link Utilization, System Utilization 9
Flows Start at the Same Time flow 1 0.9 Normalized Throughput flow 2 0.8 SCTP 0.7 TCP-SACK 0.6 0.5 0.4 0.3 0.2 0.1 0 5 10 20 Tight Link Bandwidth (Mbps) 10
TCP-SACK and SCTP flows grow together SCTP Normalized Throughput 1.5 TCP-SACK 1 0.5 100 200 300 400 500 Simulation Time (sec) 11
TCP-SACK and SCTP flows grow together SCTP Normalized Throughput TCP-SACK 1.5 1 0.5 0 60 65 70 75 80 Simulation Time (sec) 12
One Flow gives way to another Flow flow 2 SCTP gives way to TCP-SACK 0.9 Normalized Throughput flow 1 1 0.8 flow 1 2 TCPS 0.7 SCTP 1 0.6 SCTP 2 0.5 0.4 0.3 flow 2 TCP-SACK gives way to SCTP 0.9 Normalized Throughput flow 1 1 0.2 0.8 flow 1 2 SCTP 0.1 0.7 TCPS 1 0 0.6 TCPS 2 5 10 20 0.5 Tight Link Bandwidth (Mbps) 0.4 0.3 0.2 0.1 0 5 10 20 Tight Link Bandwidth (Mbps) 13
SCTP gives way to TCP-SACK 2 SCTP Normalized Throughput TCP-SACK 1.5 1 0.5 200 400 600 800 1000 1200 1400 1600 Simulation Time (sec) 14
SCTP gives way to TCP-SACK SCTP 2 TCP-SACK Normalized Throughput 1.5 1 0.5 0 75 80 85 90 95 100 105 Simulation Time (sec) 15
TCP-SACK gives way to SCTP 2 Normalized Throughput SCTP TCP-SACK 1.5 1 0.5 200 400 600 800 1000 1200 1400 1600 Simulation Time (sec) 16
TCP-SACK gives way to SCTP SCTP Normalized Throughput 1.5 TCP-SACK 1 0.5 0 75 80 85 90 Simulation Time (sec) 17
Conclusion ● Irrespective of when a flow starts and bandwidth, SCTP outperforms TCP-SACK and TCP-NewReno by 35%-41% (due to better loss recovery mechanisms) ● TCP and SCTP traffic can grow together & TCP gives a Single-homed SCTP is TCP-friendly though it way to SCTP and vice versa even for the most achieves higher throughput than TCP just as TCP- imbalanced cases SACK or TCP-Reno perform better than TCP-Tahoe 18
TCP-Friendliness of CMT* * Experimental extension to SCTP (J. Iyengar, PEL @Univ. of Delaware, 2006)
Motivation ● TCP-F is defined for end-to-end transport connections over a single-path ● J. Iyengar et. al. studied performance of CMT with the assumption of bottleneck-independent topology How does CMT behave when the tight link is shared between the CMT subflows and other TCP flows? 20
Experimental Framework 100 Mbps 100 Mbps 14 msec 14 msec 100 Mbps 2 msec ● n= 8, 16, 32, 48, 64 TCP flows from Si to Di in the background. Then, introduce either (I), (II), or (III) (I) TCP1 (A1 to B1), TCP2 (A2 to B2) (II) SCTP1 (A1 to B1), SCTP2 (A2 to B2) (III) two-homed CMT where CMT-sub1 (A1 to B1), CMT-sub2 (A2 to B2) ● RED queue @ tight link with minth= 5pks, maxth = 3*minth, wq = 0.002, maxp = 0.02, buffer size = BWxDelay ● Metrics : Per-flow Throughput, Avg. flow Throughput, Fairness Index 21
Research Questions ● What is the bandwidth share of two-homed CMT compared to two independent TCP or SCTP flows? ● What is the cost of introducing one two- homed CMT flow into the network compared to two independent TCP or SCTP flows? 22
Hypotheses (I) Introducing two TCP flows: all TCP flows get an equal share of the bandwidth. (II) Introducing two SCTP flows: SCTP flows get >= share of the bandwidth compared to TCP flows. (III) Introducing one two-homed CMT flow*: CMT flow gets >= share of the bandwidth compared to two TCP or SCTP flows * CMT shares TSN space and ACK and more resilient to losses (J. Iyengar, 2006) 23
Results 25 CMT SCTP1+SCTP2 TCP1+TCP2 20 2*bw/(n+2) Throughput (Mbps) 15 10 5 0 10 20 30 40 50 60 n (greedy TCP flows in the background) 24
Results 50 CMT SCTP1+SCTP2 2*bw/(n+2) 40 Throughput (Mbps) 30 wq= 0.002 20 10 50 0 CMT 2 4 6 8 10 12 SCTP1+SCTP2 2*bw/(n+2) n (greedy TCP flows in the background) 40 Throughput (Mbps) 30 wq=0.001(RED reacts less 20 aggressively on bursty traffic) 10 0 2 3 4 5 6 7 8 9 n (greedy TCP flows in the background) 25
Results 10 avg. w/ CMT avg. w/ SCTP1+SCTP2 avg. w/ TCP1+TCP2 8 bw/(n+2) Throughput (Mbps) 6 4 2 0 10 20 30 40 50 60 n (greedy TCP flows in the background) 26
Conclusion ● CMT <= two SCTP flows (due to bustier traffic created by CMT) ● CMT > two TCP flows (due to better loss recovery and resilience to losses b/c of TSN space and ACK information) Two-homed CMT is TCP-friendly though it achieves higher throughput than two TCP flows just as two ● CMT has AIMD-based congestion control which TCP-Reno flows would outperform two TCP-Tahoe allows other TCP flows to co-exist in the network flows 27
Discussion and the End… • Other CMT-like schemes (CP, MulTFRC, mulTCP, MPAT, PA-MulTCP, MPTCP, …) ● Criticism to TCP-Friendliness (i.e, Flow-Rate Fairness) – Cost Fairness (B. Briscoe) ● TCP-F (or another fairness criteria) should include multihoming and CMT! 28
Recommend
More recommend
