Model Assisted Path Characteristic Inference

1
Model Assisted Path Characteristic Inference

• Wei Wei
• Dept. of Computer Science
• UMass, Amherst
• weiwei_at_cs.umass.edu

2
Virtual Queuing Delay Distribution of Loss Packets

• Virtual Queuing Delay (VQD) - Would-be queuing
  delay if loss packet is not lost
• Virtual Queuing Delay Distribution (VQDD) of loss
  packets plays key role in inferring some path
  characteristics
• Is there a dominant congested link along the
  path?
• What is the smallest full queuing delay for
  routers along the path?

3
Existing Method to Infer VQDD

• Loss pair approach
• J. Liu, M. Crovella, "Using Loss Pairs to
  Discover Network Properties," in Proceedings of
  the ACM SIGCOMM Internet Measurement Workshop
  2001, CA, November 2001
• Send packet pairs
• If one of the packets is lost (loss pair), use
  the delay of the packet that is not lost to get
  VQDD
• Otherwise, throw the trace away

4
Trace Utilization of Loss Pair Approach

• p loss rate, q 1-p, Bernoulli loss
• n of loss pairs
• Average of packet to generate n loss pairs
• n/(pq)
• Trace Utilization
• U n/(n/(pq) pq lt 1/4
• the equality holds when p 0.5
• Assume we need 1,000 loss pairs to get good VQDD
  estimate, send 100 packets per second
• p 0.05, U 0.0475, 21053 packets, 211sec
• p 0.01, U 0.0099, 101010 packets, 1010sec
• p 0.001, U 0.000999, 1 million packets,
  2.8hours

5
Trace Utilization of Model Assisted Approach

• One way periodic probing
• Use all the delay and loss trace to construct a
  model and estimate VQDD
• Trace Utilization 100
• Assume we need 100 loss packets to obtain good
  VQDD estimate, send 100 packets per second
• p 0.05, 2,000 packets, 20 sec
• p 0.01, 10,000 packets, 100 sec
• p 0.001, 100,000 packets, 1000sec

6
Challenge

• Construct a Model which can estimate VQDD
  accurately
• Interpolating method?
• NO. Naïve method.
• Modified HMM which can handle missing data?
• NO. Delay is not included in the state space.
  VQDD estimation is not stable with respect to
  of parameters and initial model selection.
• HMM which can handle missing data and with both
  hidden states and observations in state space?
• YES! Good match with virtual probes. VQDD is
  stable with respect to of parameters and
  initial model selection.

7
Some Results
Two lossy links
One Lossy Link
8
On the Existence of Dominant Congested Link

• ns simulations
• Simulation setup
• Topology 3 links, link bandwidths lt 10M
• Loss rate around 5
• Loads of background traffic, cross traffic (http,
  TCP, UDP,)
• 100 percent accurate when trace is longer than
  1 minute
• Internet experiments
• UMASS to UFRJ One dominant congested link on
  Brazil size. Which link?
• UMASS to USC No dominant congested link

9
Thanks !