Nettimer: A Tool for Measuring Bottleneck Link Bandwidth

1
Nettimer A Tool for Measuring Bottleneck Link
Bandwidth

• Kevin Lai and Mary Baker
• Stanford University

2
Introduction

• Bandwidth
• Rate at which data travels along a path
• Types of bandwidth

WWW
P2P file sharing
3
Nettimer Applications

• Benchmarking
• Verify vendor specifications
• Verify ISP ? customer service agreements
• Content Adaptation
• Transcode images
• Select video stream
• Performance-based routing
• Selecting from replicated servers/proxies
• Peer-to-Peer overlay networks

4
Approach
Routers
Source
Destination

• Measure bottleneck (lowest bandwidth) link
• Passively observe at network edges
• Capture packet headers at source and/or
  destination
• No modifications to core infrastructure
• No probe packets required

5
Theory Packet Pair Property

• Bandwidth packet size / ?treceive
• Observed Bolot93
• Derived from multi-packet model LB2000
• Use existing traffic (passive measurement)
• Only two packets

6
Histogram

• Assume no downstream queueing?
• Not valid
• Assume cross traffic is not strongly correlated
  in size or arrival time
• Carter,Crovella96 Paxson97

7
Kernel Density Estimation

• Histograms not robust
• Bin size?
• Kernel Density Estimation is robust LB99

8
Passive Distributed Measurement Architecture
Distributed Packet Capture Server
Distributed Packet Capture Client
Measurement Path
Measurement Data Connection

• Servers capture packet headers and send to
  clients
• (n2-n)/2 measurement paths
• Clients run bandwidth measurement algorithms
• New algorithms only require changing client
• Clients bear compute burden
• Consume only a fraction (6) of bandwidth in use

9
Measurement Direction
RBPP
SBPP
ROPP

• Receiver Based Packet Pair (RBPP) Paxson97
• Most accurate, need to deploy at both hosts
• Sender Based Packet Pair (SBPP) Paxson97
• Away direction only, deploy only at local host
• Receiver Only Packet Pair (ROPP) LB99
• Towards direction only, deploy only at local host

10
Measurement Methodology

• Bottleneck link
• Ethernet 100Mb/s, 10Mb/s
• WaveLAN 11Mb/s, 2Mb/s
• ADSL 1.5Mbs/128Kb/s
• V.34 Modem 33.6Kb/s
• CDMA data 19.2Kb/s
• Path length 3-18 hops, Stanford, CA to Harvard,
  MA
• Data transfer direction
• Measurement times 1807 Fri, 1636 Sat, 1107
  Mon, 1839 Mon, 1200 Tue

11
Results
12
Results Agility
13
Results Agility
14
Results Agility
15
Conclusion

• Future Work
• Compute confidence in estimate
• Measure traffic load
• Bottleneck Link Bandwidth Measurement
• Minimally invasive, fast, accurate, easily
  deployable
• Nettimer tool
• GPL, Linux gt 2.2
• http//mosquitonet.stanford.edu/laik/projects/net
  timer