1. CSE 4213: Computer Networks II
Suprakash Datta
Office: CSEB 3043
Phone: ext 77875
Course page:
4/22/2019
COSC S.Datta

2. TCP congestion control – the problem
Roadmap
TCP congestion control – the problem
TCP congestion control – solutions
Performance of TCP congestion control
The future of TCP
4/22/2019
COSC S.Datta

3. TCP congestion control: the problem
Arguably the most crucial part of TCP
Lots of research effort (over 25+ years)
Performance Objectives
Link utilization
Fairness
Keep congestion down
Q: What is the right notion of fairness?
4/22/2019
COSC S.Datta

4. TCP congestion control: the problem
Nodes do not know about
other nodes
total or available capacity on any link
the number of packets in a router buffer
So, how does a node sense
congestion?
delays, losses,….
absence of congestion?
lower delays, fewer losses….
4/22/2019
COSC S.Datta

5. TCP congestion control – the problem
Roadmap
TCP congestion control – the problem
TCP congestion control – solutions
Performance of TCP congestion control
The future of TCP
4/22/2019
COSC S.Datta

6. TCP congestion control – solutions
“Packet loss indicates congestion”
Sensing congestion: timeouts, duplicate ACKs
Reacting to congestion: drastic decrease of sending rates
How/when does a sender recover from congestion?
Continuous probing
Conservative increases
Q: Formula for increase and decrease ?
4/22/2019
COSC S.Datta

7. TCP congestion control – v1
Let’s think in terms of rates:
Additive (conservative) increase, multiplicative (drastic) decrease of sending rates (AIMD)
Starting rate: High? Low?
Transmission rate
24 Kb/s
16 Kb/s
8 Kb/s
time
4/22/2019
COSC S.Datta

8. TCP congestion control – v2
TCP Tahoe (1988)
Two modes: Aggressive probing (Slow start) and careful probing (Congestion avoidance).
Slow(!) start mode: Multiplicative (aggressive) increase, multiplicative decrease of sending rates.
Congestion avoidance mode: Additive (conservative) increase, multiplicative (drastic) decrease of sending rates.
4/22/2019
COSC S.Datta

9. TCP Tahoe illustration
Assume that ssthresh = 8 14 12 10
ssthresh 8
ssthresh 6
Cwnd (in segments) 4 2
t=0
t=2
t=4
t=6
Rounds
4/22/2019
COSC S.Datta

10. TCP Reno(1990) and Fast recovery
A further improvement
TCP Reno(1990) and Fast recovery
Intuition: Distinguish between loss and 3-dup ACK
4/22/2019
COSC S.Datta

11. TCP uses windows (number of packets sent without waiting for ACKs)
Windows vs rates
TCP uses windows (number of packets sent without waiting for ACKs)
Transmission rate = W * Packet size/RTT
RTT = round trip time
Self-clocking
4/22/2019
COSC S.Datta

12. TCP congestion control – the problem
Roadmap
TCP congestion control – the problem
TCP congestion control – solutions
Performance of TCP congestion control
The future of TCP
4/22/2019
COSC S.Datta

13. Performance of TCP congestion control
Fairness
Link utilization
Algorithmic perspective?
Note: Many other TCP versions have been proposed. No one version outperforms all others in all situations
4/22/2019
COSC S.Datta

14. TCP Performance – Fairness
[Chiu and Jain, 89] Two competing LONG sessions, no arrival or departures, assume rates are real numbers. R
equal bandwidth share
loss: decrease window by factor of 2
Session 2 throughput
congestion avoidance: additive increase
Session 1 throughput R
4/22/2019
COSC S.Datta

15. Once the trajectory hits x=y line, it stays there
Lessons
Caveat: In reality, both sessions may not detect congestion at the same time
Once the trajectory hits x=y line, it stays there
For real-valued rates, reaches x=y only as t 
Will leave optimal point after reaching it (continuous probing)
Generalizes to n sessions
MIMD, MIAD, AIAD all unfair
Can be shown to hold in the presence of arrivals and departures if all sessions are long [Edmonds, Datta, Dymond, 03]
4/22/2019
COSC S.Datta

16. TCP Performance – link utilization
At least 25% capacity lost
Transmission rate C
3C/4
C/2
time
4/22/2019
COSC S.Datta

17. TCP Performance – algorithmic view
How good an algorithm is it?
i.e. Optimal? Near-optimal? ….
Very difficult question
- Does not correspond to clean theoretical models.
[Edmonds, Datta, Dymond, 03] TCP “performs well” (is competitive against a limited adversary) when each session is of a minimum specified size…
4/22/2019
COSC S.Datta

18. TCP Performance – problems
When/where does TCP not work well?
Multimedia networking
Non-smooth (sawtooth) bit rate
Best-effort: no QoS guarantees
Wireless/hybrid networks
“Packet loss indicates congestion” ?
Congestion and loss are not always correlated
Very high-speed networks
Fairness issues: parallel sessions, multiple bottlenecks, small sessions…
4/22/2019
COSC S.Datta

19. Co-operative algorithm Policing is hard to do
TCP/IP Design issues
Co-operative algorithm
Policing is hard to do
Tracing of malicious hosts/users difficult
4/22/2019
COSC S.Datta

20. A 2-minute introduction to the TCP/IP architecture The role of TCP
Roadmap
A 2-minute introduction to the TCP/IP architecture
The role of TCP
TCP congestion control – the problem
TCP congestion control – solutions
Performance of TCP congestion control
The future of TCP
4/22/2019
COSC S.Datta

21. Active queue management schemes
The Future of TCP
Active queue management schemes
Drop packets proactively to reduce congestion
RED, REM, BLUE,…
TCP for ultra-high speed networks
FastTCP, HSTCP,….
Usually use delay for being more responsive to congestion
Less draconian
TCP for wireless, hybrid networks
Treats wireless parts separately
4/22/2019
COSC S.Datta