Predicting Internet Network Distance with CoordinatesBased Approaches

1
Predicting Internet Network Distance with
Coordinates-Based Approaches

• T.S.Eugene Ng and Hui Zhang
• CMU

2
Motivation

• Emergence of new classes of large-scale
  globally-distributed network services
• Distributed content hosting services
• Overlay network multicast
• Peer-to-peer file sharing
• Prediction of network distance can greatly help
  these systems to intelligently choose their
  communication paths

3
Objective

• Devise techniques that can predict network
  distance accurately, scalably, and in a timely
  fashion
• Three techniques
• IDMaps
• Triangulated Heuristic
• GNP (proposed in this paper)

4
IDMaps

• Special HOPS servers maintain a virtual topology
  map of the Internet consisting of end hosts and
  Tracers
• Distance of host A and B is estimated as Dabc

Tracer 1
a
b
A
c
B
Tracer 2
5
Triangulated Heuristic

• Select N nodes in the network to be base nodes
• A node H is assigned coordinates (N-tuples) which
  are the distances between H and the N base nodes
• Then the distance between H1 and H2 is bounded
  below by
• bounded above by
• U, L, or (UL)/2 can be distance estimation

6
Trangulated Heuristic
Base nodes
A
B
7
Global Network Positioning(GNP)

• Notations
• S model the Internet as a particular geometric
  space
• the coordinates of a host H is S
• the distance function that operates
  on the coordinates
• the
  computed distance between host H1 and H2
• measured distance

8
GNP_Landmark Operations

• Use N hosts , known as Landmarks to
  provide a set of reference coordinates
• Using ICMP ping messages to produce the bottom
  half of the NxN distance matrix
• Compute the coordinates of the Landmarks by
  minimizing
• where

9
Landmark operation
10
Ordinary Host Operation

• Ordinary host derives its own coordinates by
  using the coordinates of the landmarks
• The host coordinates can be obtained by
  minimizing

11
Ordinary Host Operation
12
Special issues

• One common point of the three methods the need
  for some infrastructure nodes(i-nodes) tracers
  for IDMaps, base nodes of the triangulated
  heurisic, Landmarks of GNP.
• How to choose N and how to deploy them (open
  question)?

13
How to choose N?

• Maximum separation maximize the total
  inter-chosen-probe distances
• N-medians minimize the total distance from each
  not-chosen probe to its nearest chosen probe.
• N-cluster-medians form N clusters of probes and
  then choose the median of each cluster as the
  i-nodes.

14
Performance evaluation

• I-nodes chosen from 19 hosts distributed around
  the world, called probes
• A bunch of IP addresses, called targets
• Performance metric directional relative error

15
Relative error comparison
16
Number of N
17
Error Measurement Function

• Normalized error measure
• Logarithmic error measure

18
Predicting TCP Throughput From Non-invasive
Network Sampling

• M.Goyal, R. Guerin, R. Rajan
• OSU

19
Motivation

• The need for service verification and quality
  monitoring is increasing
• Convert network observables (packet drop ratio)
  into representative user and application relevant
  performance metrics (throughput)
• Existing models (Amherst model) cannot give
  accurate throughput prediction

20
Failure of Amherst model
21
Why?

• In Amherst model
• The probability p is the first packet loss in a
  cwnd (can not accurately be estimated).
• The packets following the first lost packet in
  the cwnd are all lost
• The initial congestion window is always half of
  the window size of the ending window size of the
  previous epoch

22
Modifications to Amherst Model

• Random loss model is introduced
• The probability of retransmission time-out

23
Modifications(cont.)

• Initial window size is the
  probability that equals

24
Throughput

• Eventually the steady state throughput of a TCP
  flow is

25
Performance Evaluation
26
Throughput
27
Throughput