Slide 1 18-Feb-02

1
Progression and propagation of internet congestion

• Jim Gast
• University of Wisconsin / Madison
• Feb 19, 2002

2
When are losses likely?

• When queuing delay is high?
• When queuing delay is rising fast?
• When queuing delay levels off?
• Randomly?

Queuing delay component in milliseconds
Time
3
Research Question

• What are the characteristics of congestion in the
  Internet?
• Duration
• Propagation /location
• Complexity
• Frequency / periodicity
• Implications

4
Challenges

• Accurately recreating congestion in the lab
• multiple bottlenecks, appropriate length paths
• Visualizing congestion to gain understanding
• choosing the right level of detail (router vs.
  AS)
• Inferring congestion from data we can
  realistically gather
• passive vs. active
• packet headers vs. flow level

5
Your fair share
300
800
400
1100
800
1200
400
Assume 4 flows through 3 links. Each flow can
support speeds up to 1000 packets per second.
6
Your fair share
At time t1, the blue flow ends.
300
800
0
400
1100
800
1200
400
800?
When one of the flows ends, there is reduced
congestion at the link with capacity 800. After
t1, the black flow slowly increases toward 800
pps.
7
Your fair share
300
600?
800
0
400
1100
800
1200
800
600?
Increased flow of black packets causes congestion
at the middle link. It responds by randomly
dropping brown and black packets until the sum of
brown and black drops to 1200.
8
Your fair share
300
500?
600?
800
0
400
1100
800
1200
800
600?
Congestion at the right link disappears.
Decreased flow of brown packets leaves more room
for green so greens usage slowly rises toward
500 pps.
9
Queue Depths
Full
Empty

Demand 70 74 78 96 100 104 70
Assume many flows converge at a particular link
of capacity 100 packets per second. When
aggregate demand exceeds 100,the queue starts to
build. When the queue fills, the router
dropspackets. Sometime later the senders learn
about the loss, slowdown, and the queue length
plummets.
10
Time Frames
Full

Empty
1
3
2
1 No loss. Duration based on multiplexing
factor, connection durations, RTT mix.
Typ? 30 seconds to 30 minutes? 2 Increasing
queuing delay. Duration based on queue size,
multiplexing, conn duration. Typ? 3 seconds to
10 seconds? 3 Congestion event. Duration
depends on shortest RTT plus three triple
dup acks. Typ? 100 millisec to 300 millisec.
11
Cascaded delays and congestion

• Cumulative queuing delay will show brief periods
  ofelevated delay followed by a congestion event
• Events will be spaced at seemingly random
  times,but with a period equal to the product of
  the periodsof the individual sites.

12
Cascaded delays and risk of losses
(Red lines risk of loss from links A or B)
Queue Depth Link A
Time

Link B

Link C

Link D

Total
13
Delay Count
Shelf depends on theslope of building
andemptying regimes
Count
Sampled at random intervals
Height dependson duration ofcongestion event
Delay
Delay
Time
underfed
congested
Routers dont tell us queue depth each
millisecond. We have to discover it by probing.
building
emptying
14
Additional Delay
Count
Delay
Delay
Time
Queues arent always completely empty. If light
cross traffic arrives at Poisson intervals, the
delay counts will show an exponential
distribution.
15
Sum of Delays
Count

?
Delay
Router 1
Router 2
16
Surveyor Data

• One-way delay measurements
• GPS time accuracy
• Feb 1999 to Aug 2000

17
Full Queue?

• Zoom

18
Level Shift
19
Conceptual (unproven) Simulation
300
400
800
1100
800
1200
400
20
Next Steps

• Wavelet Analysis
• External Measurements
• NIMI
• Surveyor
• WAWM
• GIMI

21
Implications

• Improved tuning and performance of existing
  protocols
• RED
• ECN
• Fewer surprises when the next killer app hits
  campus
• Basis for cross-protocol next generation
  congestion control primitives
• Lossless congestion control
• Reduced delay, reduced jitter, reduced
  retransmissions when sharing
• Improved Internet fairness, predictability, and
  scalability

22
Questions?
23
Simple bottleneck - pacing losses
24
Multiple Congestion
Disruptive Losses
Disruptive Loss
Desired Signal
25
Analysis Example
Disruptive Losses
26
Shared Congestion Suspicion
Total of congestion reports from all of the
connections that pass through this site
Component contributed by the connection in the
prior slide