Network Available Bandwidth Measurements

1
Network Available Bandwidth Measurements

• XiaoLong HE

2
Outline of the Talk

• Problems
• Related Work
• Traffic Design
• Our Approach (SimProbe)
• Analysis of Simulation Results
• SMRT Implementation
• Lessons Learned
• Future Work

3
Why measurement?

• Measurement is critical to network control.
• Measurement for selecting server/ISP
• Measurement for verifying network configuration
• Measurement for designing real-time applications
• Measurement for load balancing in WAN
• Measurement for accounting

4
Taxonomy of Internet Network Measurement
Approaches

• Sender-based vs. Receiver-based
• Passive Watch vs. Active Probe
• Layer of Protocol used

5
Measurement Problems

• Route Fluttering
• Packet Loss
• Out of Order packet delivery
• Clocks Synchronization
• Routers reject ICMP due to denial of series
  attacks.

6
Bottleneck Bandwidth

• Bottleneck bandwidth the maximal rate at which
  data can be sent along a connection, determined
  by the slowest element in the entire chain.

7
Available Bandwidth

• Available Bandwidthhow fast can the data in fact
  be transmitted over the connection.
• PacketSize/(PacketSizeCompetitorPacketSize)
  Bottleneck Bandwidth

8
Related Works

• Bprobe/Cprobe - Robert L. Carter and Mark E.
  Crovella, Boston University, 1996
• Ping - M. Muuss, U.S.Army Ballistic Research
  Laboratory, 1983
• Traceroute - Van Jacobson, Berkeley, Univeristy
  of California, 1997
• Treno - M. Mathis and J. Mahdavi, Proceedings of
  INET96, Montreal, 1996

9
Bprobe/Cprobe

• Bottleneck Bandwidth (PacketSize/Time)
• Available Bandwidth (S2Sn)/(Tn-T1)

10
SimProbe Simulator

• To carry the thesis out, we build a Java-based
  discrete event simulator.
• Swing, Java2D and AWT packets for GUI design
• Thread for comparing different probing methods
• Port it to Unix (without GUI) for long time
  simulation

11
How traffic is generated in SimProbe Simulator
12
Traffic Patterns supported by SimProbe

• Flat/Static
• Slope
• Web (modifed code by Arthur Blais)
• Time between requests Pareto Distribution
• Number of embedded references Pareto
  Distribution
• Time between references Weibull Distribution
• File Size Body (Lognormal Distribution) Tail
  (Pareto Distribution)

13
Network Topology Used in This Simulation Study
14
Our Approaches

• Sending as few packets as possible. Using round
  trip time measurements for network available
  bandwidth.
• Trip Time Measurement FormulaTripTimeMeasuredAv
  ailableBandwidth MinimalTripTimeMaximalAvailabl
  eBandwidth
• Round Trip Time Measurement
• SMRT
• MMRT
• Unidirectional Trip Time Measurement
• SMUT
• MMUT
• Time Gap between packets
• ACB (Adaptive Cprobe)

15
SMRT(Single Msg Round Trip Time Measurement)

• Using multiple packets to find minimum of the
  round trip time of a connection
• Using Bprobe to get the maximal available
  bandwidth of a connection
• Sending single probing packet
• Using round trip time to estimate the available
  bandwidth of the connection

16
MMRT

• Use Bprobe to get the maximal available bandwidth
  of a connection
• Using multiple packets to find minimum of the
  round trip time of a connection
• Sending multiple packets
• Round trip time for estimating ABW

17
SMUT

• Use Bprobe to get the maximal available bandwidth
  of a connection
• Using multiple packets to find minimum of the
  unidirectional trip time of a connection
• Sending single probing packet
• Using unidirectional trip time for estimating
  available bandwidth

18
MMUT

• Use Bprobe to get the maximal available bandwidth
  of a connection
• Using multiple packets to find minimum of the
  unidirectional trip time of a connection
• Sending multiple probing packets
• Using unidirectional trip time for estimating
  available bandwidth

19
Simulation result analysis

• Difference MetricAvailableBandwidth -
  MeasuredAvailableBandwidth AvailableBandwidth
• AvailableBandwidth PacketSize/(PacketSizeQueui
  ngSize) Bottleneck Bandwidth
• MeasuredAvailableBandwidth (MinimalTripTimeMa
  ximalAvailableBandwidth)/ TripTime
• DelayTimeTripTime - MinimalTripTime

20
(No Transcript)
21
(No Transcript)
22
SMRT Random Characteristics

• Problem Dependent on the arrival time of probing
  packet on bottleneck link
• Three Scenarios
• 1) no queuing time, yields bottleneck available
  bandwidth.
• 2) yields the lower bound of bottleneck available
  bandwidth
• 3) yields average bottleneck available bandwidth

23
Impact of Packet Size on SMRT

• DelayTime rises when probing packets size
  increases.
• Smaller packet size gets worse results
• RoundTripTime MinimalRoundTripTime DelayTime
• Examples
• T1 / (T1?T)
• T2 / (T2?T)

24
Impact of Link Length on SMRT

• The result accuracy is worse with the length
  increasing
• Delay Time dominates RoundTrip Time at shorter
  link length
• Propagation time dominates RoundTrip Time at
  longer link length

25
Impact of Router Speed on SMRT

• No apparent difference for 1GBps or 10GBps
• Router Speed affects the time gap between packets
• Slow router speed enlarges the time gap between
  packets
• Router Speed must be faster than link speed

26
Impact of Bottleneck Link Bandwidth on SMRT

• SMRT results follow the available bandwidth trend
• SMRT performance is similar to the performance
  trend by SMRT performance with 10MBps bottleneck
  link bandwidth

27
MMRT

• Try to alleviate the random characteristic of
  SMRT
• Small time gap between probing packets may
  experience big traffic difference in Slope
  traffic pattern
• Obvious improvement on SMRT performance in web
  traffic

28
SMUT

• Measure the available bandwidth on unidirectional
  route.
• SMUT has the same random characteristics as
  SMRT.

29
MMUT

• Try to alleviate the random characteristic of
  SMUT
• Small time gap between probing packets may
  experience big traffic difference with slope
  traffic pattern
• MMUT has Better performance than SMUT with web
  traffic pattern

30
ACB

• Using the last measured available bandwidth
• Send two packets once a time
• Reprobing in Cprobe if time gap between packets
  is inside of a range
• Not better than Cprobe

31
Summary of Simulation Analysis
32
Implementation of SMRT

• SMRT is implemented in C
• Maintain a host file, which specifies the list of
  sites to be probed
• Bprobe is used to measure the largest bottleneck
  available bandwidth
• Ping is used to measure the smallest round trip
  time of a connection
• Using Bprobe and Ping to collect the history
  information for each host in the file, at offpeak
  and over a long period
• Probed host must be contained in the host file
• Round trip time of the probing message is used to
  measure the instant available bandwidth

33
Implementation results of SMRT

• We probed a list of mirror sites for popular
  software packages such as perl, linux and
  netscape

34
Implementation results of SMRT

• Orst is good choice for us to download software

35
Implementation results of SMRT

• Orst gives us relative short round trip time

36
Lessons Learned

• Impact of router speed on traffic generation
• Problem When router speed is lower than the
  bottleneck link speed, we can not get expected
  traffic?
• Solution Make sure the router speed in our
  connection is greater than the bottleneck link
  speed, especially for the large bottleneck link
  bandwidth

37
Lessons Learned (Cont)

• Cprobe Limitations
• Probing returned packets interfere with later
  probing packets
• Queuing time is not the only results from
  connection competition
• Examined factors
• Router Speed
• Hop Number
• Message Size
• Message Number

38
(No Transcript)
39
(No Transcript)
40
Lessons Learned (Cont)

• OOP design in Java
• Problem Chart can not be updated when we call
  the chart in a thread directly.
• Solution Using SwingUtilities.invokeLater Class
• Double data type precision problem
• Problem The round trip time with traffic pattern
  is even shorter than the time without any traffic
• Reason machine double type data has a precision
  problem in representing a value exactly
• Solution Because difference is small, so ignore
  it

41
Future Works

• How to measure available bandwidth of a host
  whose history information is not available.
• How to find a better available bandwidth
  estimation formula to improve the accuracy of
  SMRT
• Convert MMRT, SMUT and MMUT simulation code from
  Java to C
• More traffic route to test the reliability of
  SMRT and other methods