Download presentation
Presentation is loading. Please wait.
Published byTrevor Allison Modified over 6 years ago
1
TCP Vegas: New Techniques for Congestion Detection and Avoidance
Team 1: Cheng-Lin Tsao and Ruhull Alam Bhuiyan
2
Introduction TCP Reno TCP Vegas Congestion control
Coarse timeout/retransmission Need of creating losses TCP Vegas Congestion avoidance Accurate timeout, timely retransmission Congestion detection 5 techniques 40-70% better throughput, 1/2-1/5 losses
3
Background and Motivation
Knee and cliff Congestion avoidance Operate around knee Congestion control Operate on the left of cliff Motivation Limit of congestion control Coarse timeout/retransmission
4
Technique 1, 2 More accurate RTT calculation
Reno: coarse-grained timer (500 ms) 1100ms to timeout and retransmit Vegas: read/record clock on sending segments/receiving ACKs <300ms to timeout and retransmit New retransmission mechanism Reno: 3 duplicate ACKs Vegas: check timeout when receiving duplicate ACK 1st or 2nd non-duplicate ACK after retransmission
5
Technique 3, 4 Modified window sizing on congestion Spike suppression
Reduce cwnd only due to losses at the current sending rate Reduce cwnd only if retransmitted segment was sent after last decrease Spike suppression Multiple segments transmitted back-to-back Large cumulative ACKs or ACK compression Space segments evenly to prevent buffer overflow
6
Technique 5 Congestion detection Modified slow start Concept Algorithm
Expected = WindowSize / BaseRTT Modified slow start
7
Performance Evaluation Model
Simulator University of Arizona’s x-kernel framework Full protocol stack Protocol is modeled by actual C code Traffic generation with tcplib Tracing facility Network configuration
8
TCP Reno with no other traffic
Slow start Congestion control (dashed) slow-start threshold (dark gray) flow control window Detect loss, decrease cwnd Decrease cwnd to half and set ssthresh (thin) actual # bytes in transit (light gray) congestion control window Linearly increase cwnd (solid vertical) time when a retransmitted segment originally sent Exponentially increase cwnd in every RTT Incur packet loss Incur packet loss sending rate Overrun router buffer Overrun router buffer Use more and more buffer
9
TCP Vegas with no other traffic
Slow start Congestion avoidance Increase cwnd Enter linear increase/decrease mode Exponentially increase cwnd in every other RTT Steady sending rate Actual > α (gray) Expected throughput congestion decision (dashed) thresholds of throughput (solid) Actual throughput Actual throughput < γ β < Actual < α α < queue < β Reduce sending rate before buffer overrun Queue < α
10
Results (1) How 2 TCP connections interfere with each other
4 possible combinations Reno/Vegas: a 300KB transfer using Reno contained within a 1MB transfer using Vegas Reno’s throughput unchanged when competing with Vegas Vegas doesn’t adversely affect Reno’s throughput
11
Results (2) Performance of TCP with background traffic
Two versions of Vegas Vegas-1,3: α = 1 and β = 3 Improvement of Vegas over Reno More than 50% improvement on throughput Half the losses
12
Results (3) Measurements of TCP over the Internet
Connection between UA and NIH 17 hops 37-42% more throughput, 51-61% retransmissions
13
Related Work Proactive congestion detection Wang and Crowcroft, DUAL
Based on understanding of network changes as it approaches congestion Wang and Crowcroft, DUAL compare RTT to (min + max)/2 Jain, CARD compare RTT and congestion window size Wan and Crowcroft, Tri-S compare throughput to window 1 segment smaller Keshav, Packet-pair rate probing estimate available bandwidth with two segments Vegas measures and controls throughput amount of extra data a connection has in transit
14
Critique and Discussions (1)
Unfairness problem α and β are thresholds of extra data a connection has in transit Including data of other connections Unfairness among connections with different hop counts Ex. R1~R4 have extra data of 1 segment, H1~H4 measure 2 extra data, H5 measures 4 extra data. H5 always reduces its sending rate while others remain the same. H1 H2 H3 H4 H5 R1 R2 R3 R4 H6
15
Critique and Discussions (2)
Linear or exponential decrease Available bandwidth for each connection may drop exponentially due to critical changes, such as node failure R2 R1 R4 R3
16
Critique and Discussions (3)
Influence from application layer Multimedia: steady data rate HTTP: slower slow start More accurate measurement and control Why Expected throughput > available bandwidth? Invalid BaseRTT when route changes Optimal values of α, β, and γ? UDP with congestion avoidance Vegas decouples reliability and congestion control link
17
Summary and Conclusions
Vegas is TCP with congestion avoidance It controls extra data the connection occupies in the network It has more accurate RTT and timely retransmission mechanism Experiments and simulation both show 40-70% better throughput and half retransmission, compared to Reno
18
Expected throughput over-estimated
Expected throughput 240KBps Available bandwidth only 200KBps back
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.