Measuring End-to-end Bandwidth with Iperf using Web100

1
Measuring End-to-end Bandwidth with Iperf using
Web100

• Presented by Warren Matthews (SLAC)
• on behalf of
• Ajay Tirumala (U of Illinois), Les Cottrell
  (SLAC) and Tom Dunigan (ORNL).

2
Jim said

• Comments and criticisms of how easy it is to
  download, patch, etc.
• Comment on the completeness of the documentation
  we have available.

3
Not Really Web100 Issues

• Unix/Linux machines on the SLAC network are
  heavily taylored. Heavy reliance on AFS.
• Problems with running customized kernels in
  taylored (considered secure) environment
• Developers very responsive

4
SLAC and Web100

• Future of HENP depends on network performance
• Genetics, Astronomy, VoD
• Ability to monitor and tune TCP is essential
• Web100/Net100 very useful tool

5
Abstract

• Iperf considered accurate but intrusive
• Use web100 to detect end of slow start
• Measurement within 10
• Save 94 time, 92 traffic

6
Slow Start (Theory)
7
Standard Iperf

• Tool of choice for many network admins
• Errors due to long slow start
• Typically run for 20 seconds (default is 10)

8
Typical Results
9
Non-Typical Results
10
Modified Iperf

• Modify Iperf client
• Web100 required on client only
• Paper describes algorithm to detect end of slow
  start
• Included in release

11
Comparison
12
Further Work

• Test with new TCP stacks
• Expect will work better
• Test with faster links
• GigE
• 100 ms latency
• Correlate congestion signals, loss and
  retransmission information with bandwidth
  achieved.

13
Further Information

• Paper presented at PAM