Diagnosing SpatioTemporal Internet Congestion Properties

1
Diagnosing Spatio-Temporal Internet Congestion
Properties

• Leiwen Deng
• Aleksandar Kuzmanovic
• EECS Department
• Northwestern University

http//networks.cs.northwestern.edu
2
Problem

• Detect congestion events on an end-to-end path
  and reveal their spatio-temporal properties
• Where they happen (edge, core, intra-AS,
  inter-AS)?
• How long they last / frequently occur?

S
D
3
Why Do We Care?

• Fault diagnosis
• Advanced congestion control
• Overlay design
• Distributed monitoring systems
• We want to know!

S
D
4
Challenges

• Congestion events relatively infrequent
• Measure queuing delay instead of Ploss
• No/low support from the network
• Combine e2e with probes to intermediate nodes
• Path asymmetry
• Measurements still possible via measurable
  pairs

5
Outline

• Methodology
• Implementation (Pong)
• Validation
• Internet measurements

6
Methodology Highlights

• Coordinated probing
• Send 4, 3, or 2 packets from two endpoints
• Quality of Measurability (QoM)
• Able to deterministically detect its own
  inaccuracy
• Self-adaptivity
• Switch among different probing schemes based on
  QoM and path properties

7
Coordinated Probing
Probe
S
D
f probe
b probe
s probe
d probe
,
,
,
4-p probing a symmetric path scenario
8
Coordinated Probing
Probe
?f
?d
S
D
?s
?b
?fs
?fd
9
Locating Congestion Points
1. Probe Scheduling
S
D
Sequentially probe (4-p) nodes on the path
10
Locating Congestion Points
2. Switch Point Approach
S
D
Correlate probes to neighboring nodes
11
Tracing Congestion Status
S
D
Link 1 (Located Congestion Point)
Congestion Status
Link 1
Time
Reuse probes sent to un-congested routers
12
Outline

• Methodology
• Coordinated probing
• Switch point approach
• Path asymmetry
• Validation
• Internet measurements

13
Measurable Pairs
4-p probing scenario
14
Quality of Measurability
15
Demoted Probing Schemes
16
Tuning Probing Techniques

• Objective
• Tune probing techniques based on QoM

Definition of QoM
Condition
Probing technique
4-p probing Fsd probing Fsb probing 2-p probing
?f ?b ?s ?d ?f ?s ?d ?s ?f ?b unconditional
(Last resort)
17
Adaptive Pairing

• How to pair s and d probes?
• A non-trivial task
• A single s probe can
• have a number of
• complementary d
• probes
• Our approach
• Priority given to
• d probes that are
• less frequently tried
• achieve longer durations (when used)
• have larger average QoM

s
Congestion
D
S
d
b
Measurable Pair
Complementary d probe
18
Outline

• Methodology
• Implementation (Pong)
• Validation
• Internet measurements

19
Validation

• Simulations (ns-2)
• Emulab
• Self-consistency validation in the Internet

20
Experimental Setup

• Topology
• 12 nodes, 11 links
• Link 100 Mbps, 2ms
• Cross Traffic
• TCP cross-traffic, 50 time on 50 off
• Multiple bottlenecks built simultaneously

21
Evaluation
Before adding backward bottlenecks
22
Evaluation
After adding backward bottlenecks
23
Evaluation
Before adding two more forward bottlenecks
24
Evaluation
After adding two more forward bottlenecks
25
Summary

• Before probe correlations
• Pong exhibits slight bias in detecting congestion
  locations
• After probe correlations
• No bias in detecting congestion locations
• Slight bias in determining congestion frequency
• Affected only by the distance between congested
  points
• Probing techniques
• 4-p, fsd, and fsb high accuracy
• 2-p capable of accurately locating a single
  congestion point

26
Outline

• Methodology
• Implementation (Pong)
• Validation
• Internet measurements

27
Experiments

• Two independent large-scale Internet measurements
  using over 400 hosts
• The first experiment measures 23,000 paths
  within 8.5 days
• The second experiment measures 20,000 paths
  within 7 days
• Measure each path for 1 hour
• Results from the 2 measurements fully consistent

28
Results

• Edge vs. core
• Edge more frequently congested than the core 4.5
  times on average
• Intra-AS vs. Inter-AS
• Edge Intra-AS gt Inter-AS
• Core Intra-AS lt Inter-AS
• Time domain
• Edges congestion events clustered in time
• Core congestion events dispersed in time
• Links vs. Paths
• Links 12 congested, 3 considerably
• Paths 20 considerably congested

29
Multiple Congested Points

• Probability to observe multiple congested points
  on an end-to-end path
• Grows as a power function of interval length
• Decays exponentially with the number of congested
  points

30
Conclusions

• Spatio-temporal Internet congestion properties
• New methodology
• Coordinated probing
• Detect its own inaccuracy
• Self adaptive to path properties
• Handles path asymmetries
• Implemented, deployed, evaluated, measured
• High accuracy in both spatial and temporal
  domains
• Future work
• Triggered monitoring system to learn more