Bottleneck Bandwidth Estimation

1
Bottleneck Bandwidth Estimation

• Instructor Dr. Aggarwal
• Present by Jason Wei

2
Outline

• Introduction of bandwidth estimation
• Related Work
• Our algorithm and experiment
• Summary

3
Introduction

• A high-end user's connection speed grows by 50
  per year 1
• Monitoring the performance of the network,
  especially the network bandwidth, still is a
  challenging research topic.

4
Introduction (Contd)

• What is network bandwidth?
• The bandwidth of a network, also called
  throughput, is given by the number of bits that
  can be transmitted over the network in a certain
  period of time.
• For example, the bandwidth of a STS-1 network is
  51.840 Mbps, which means that the transfer
  ability of the network is 51.840 million bits per
  second.

5
Introduction (Contd)

• Bottleneck link bandwidth
• If there is no other traffic in the path,
  bottleneck link bandwidth is the maximum
  throughput that the path can provide from source
  to destination. It is limited by the bottleneck
  links underlying capacity.
• Available bandwidth
• Available bandwidth is the maximum rate that the
  path can provide to a flow without reducing the
  rate of the cross traffic.

6
Introduction (Contd)

• Why measure bandwidth?
• Network users benefit from knowing the bandwidth.
  
• Measuring bandwidth is the key to congestion
  control and QoS (Qualities of Service).
• Dynamical server selection
• It is important to know bandwidth for mobile
  computing.

7
Relation work

• Bottleneck link bandwidth measurement can be
  divided into two categories
• Single-packet algorithms
• packet-pair algorithms

8
Relation work (Contd)

• Single-packet algorithm
• Use time-to-live (TTL) field in IP packet. TTL is
  used to monitor and remove the packets.
• Once TTL is decremented to zero, the packet will
  be discarded and the router will send back an
  ICMP time exceeded packet to the original sender
• By sending out a series of probe packets with
  different values of TTL. Bandwidth and latency of
  each link can be calculated by using statistics
  to analyze the time when error packets are
  received.

9
Relation work (Contd)

• Single-packet algorithm

RTT q1 (lat1 size/bwn1) q2 forward
q3 (lat2error_size/bwn2) q4 2
10
Relation work (Contd)

• Packet-Pair algorithm 3

Bandwidth Size/Q b
11
Relation work (Contd)

• Cartouche algorithm4
• Harfoush, Bestavros and Byers presented Cartouche
  algorithm.
• Cartouche probe apm pqr-1pm can measure the
  bottleneck bandwidth of targeted path segments
  (i,j).
• s(p)gts(m)s(q),D(a)i, D(p)D(q)j, D(m)A

R1
Ri-1
R i
R j
Rn
Rj1
Sender
Receiver
12
Relation work (Contd)
Before the target path segment
13
Relation work (Contd)
After the target path segment
14
Our algorithm

• The problems of Cartouche algorithm
• Need to install software on receiver side.
• If the result will be inaccurate when the cross
  traffic becomes high.

Ra
Ra1
Ri
Rj
Rb
Rb1
Sender
Receiver
15
Our algorithm (Contd)

• Two steps
• First, use single-packet algorithm to check the
  network structure, get position of the bottleneck
  link Li.

Li
Ra
Ra1
Ri
Ri1
Rb
Rb1
Receiver
Sender
16
Our algorithm (Contd)

• Second step
• Send (Hdmrd)s probes to measure the bandwidth
  of bottleneck link.
• Size(H)gtgtSize(d)Size(m),
• Dest(H)Li, Dest(d) Li1 Dest(m)Receiver

d m m d H d m m d H
17
Our algorithm (Contd)
Before probes reach bottleneck link
Ra
Ra1
Ri
Sender
d m m d H
d m m d H

• Size(H)gtgtSize(d)Size(m)
• If there is cross traffic between the sender and
  Ri , because of s(H )gtgts(d), the interval between
  d still should be same.

18
Our algorithm (Contd)
When the probes reach bottleneck link
d m m d
d m m d H
m m
Ri
Ri1
Receiver
H
ACK
d d
ACK

• Dest(H)Li, Dest(d) Li1 Dest(m)Receiver
• Only the ICMP of packet d will be return from
  Router i1

19
Our algorithm (Contd)
After the probes reach bottleneck link, ICMP
packets come back
Ra
Ra1
Ri
Sender
d d
d d
ICMP
ICMP

• If there is no cross traffic, the interval of d
  can be keep.
• If the interval of d is large enough, the time
  they pass the link with cross traffic should be
  same.

20
Our algorithm (Contd)

• Send several times of (Hdmrd)
• Record the interval of d and analyze them to get
  the interval Idd. (The minimum value of them may
  be the value we need)
• BandwidthSize(d)(r1)/ Idd

21
The improvement of our algorithm

• Compared with single-packet algorithm
• Both use ACK, so we do not need to install
  software on receiver side.
• We just use single-packet algorithm to check the
  network structure (the position of the network
  bottleneck link)
• Our algorithm is accurate than single-packet
  algorithm in the case that cross traffic exists.

22
The improvement of our algorithm (Contd)

• Compared with Cartouche algorithm
• We use ACK. So we do not need to install software
  on receiver side.

23
Verify our algorithm

• Current work Make my own network simulator to
  verify our algorithm. (partly finished)

24
Verify our algorithm (Contd)

• Next Step
• Use Network Simulator 5to demo the algorithm.
• If possible, set up a real network with routers,
  computers, and cross traffic, implement the
  algorithm.

25
Reference

• 1. Nielsen's Law of Internet Bandwidth
• http//www.useit.com/alertbox/980405.html
• 2. Using pathchar to Estimate Internet Link
  Characteristics
• Allen B. Downey ACM SIGCOMM '99 Pages 241-250
• 3. Congestion Avoidance and Control
• Van Jacobson, In Proceedings of ACM SIGCOMM98,
  Pages 314329
• 4. Measuring Bottleneck Bandwidth of Targeted
  Path Segments
• Khaled Harfoush, Azer Bestavros, John Byers
  Boston University, 2001
• 5. Network Simulator (NS), version 2
• http//www.isi.edu/nsnam/ns/

26
Summary

• Introduce the basic network bandwidth
  algorithms.(Single packet, packet pair)
• Cartouche algorithm and problems
• Our algorithm