Network Tomography based Unresponsive Flow Detection and Control

1
Network Tomography based Unresponsive Flow
Detection and Control

• Authors
• Ahsan Habib, Bharat Bhragava
• habib,bb_at_cs.purdue.edu
• Presenter
• Mohamed M. Hefeeda
• Department of Computer Sciences
• Purdue University
• Support NSF, CERIAS, IBM

2
Motivation

• Efficient resource management by utilizing wasted
  resources
• Adaptive flows do not starve due to unresponsive
  flows
• Coordinated congestion control by propagating
  congestion information to upstream domains

3
Example

• Unresponsive flows waste resources by taking
  their share of the upstream links and dropping
  packets at downstream links are congested

4
Related Work

• Congestion collapse from undelivered packets
  Flyod et al., TON 99
• Network Border Patrol Albuquerque et al.,
  INFOCOM 00
• Edge routers periodically poll cores Chow et
  al., Internet draft 00
• Direct Congestion Control Scheme Wu et al.,
  Internet draft 00
• Loss of high class packet means congestion
• Core-assisted Congestion Control Habib, Bhargava
  PDCS 01

5
Network Tomography

• Network tomography uses correlations among
  end-to-end measurements to infer per-link
  characteristics.
• Back-to-back packets experience similar
  congestion in a queue with a high probability
  Duffield et al., INFOCOM 01

• Receiver observes the probes and correlates them
  for loss inference
• For general tree? Send stripe from root to every
  order-pair of leaves

6
Tomography-based Congestion Control (TCC)

• Only edge to edge measurements are used to detect
  and control unresponsive flows

7
TCC- Detection

• 1. Measure Delay
• Ingress routers sample user traffic
• The user packet headers are copied to probe
  edge-to-edge path for delay
• Exponential moving weighting average is computed
  with more weight to the recent history and less
  weight to the current sample
• If probed delay is higher than a specified
  threshold, a path is suspected to be congested

8
TCC- Detection (Contd)

• 2. Measure Loss
• A loss probing tree is generated with a set of
  paths that have high delay.
• The tree is probed to infer loss ratio of each
  individual link of the suspected paths
• Need to know
• Topology
• Senders
• Receivers

9
TCC- Detection (Contd)

• 3. Identify egress routers
• Through which suspected flows are leaving the
  domain. The links with high losses are feeding
  flows to these routers
• 4. Identify misbehaving flows
• These are determined with the rate at which
  suspected flows are entering into and leaving
  from a domain

10
TCC- Control

• We know the misbehaving flows from detection
• The rate of suspected flows are adjusted based on
  
• Loss ratio and
• Change of loss ratio with time

11
Experiments Evaluation methodology

• Simulation using ns-2
• Use parameter settings (queue, traffic, ) from
  reference work
• Input Parameters
• We vary RTT, number of flows, and life time of
  flows
• Output Parameters
• Measure delay, loss ratio, throughput

Topology
12
Delay Measurements

• End-to-end delay is high due to excessive flows
• With control the delay goes down

End to-End Delay (Sec)
Time (Sec)
13
Loss inference Validation

• Three different experiments
• Actual loss is close to infer loss
• Converges within 10-15 sec

Inferred Loss
Actual Loss
14
Flow Control
Bandwidth (Mbps)
TCP congestion window
Time (Sec)
Time (Sec)
Flow control mechanism increases the bandwidth of
adaptive flows by consuming bandwidth wasted by
unresponsive flows
15
Flow Control

• Loss ratio of an unresponsive flows with and
  without flow control
• Goes down sharply with time
• Converges to a low specified value

Loss Ratio
Time (Sec)
16
Flow Aggregation

• 6-10 aggregate flows of each type
• 10-100 micro flows per aggregate
• Works fine even more and more flows misbehave

Bandwidth (Mbps)
Number of flows
17
Conclusion

• A new way to detect and control unresponsive
  flows
• No involvement of core routers
• Scalable
• Easy to deploy
• Low overhead. Probe traffic less than 0.015 of
  the link capacity (OC3)