Backbone Performance Analysis Techniques

Steve Corbató <corbato@cac.washington.edu>
Networks and Distributed Computing
University of Washington
4545 15th Avenue NE
Seattle, Washington 98105-4527, USA
http://weber.u.washington.edu/~corbato/isoctalk/

Talk outline

• Motivations for understanding real-time performance
• Data collection methodology and a few caveats
• Analysis techniques
• Future work

Motivations

• Network performance
  • Detecting bottlenecks--upstream and downstream
  • Capacity planning
• Phenomenology of aggregate flows
  • Impact of potential TCP and HTTP modifications
  • Congestion collapse
  • Self-similar behavior?
• Tracking intensive real-time demand
  • Medical imaging, WWW graphics, video applications
• Comparison with long-term 15-minute data sets
  • Intermittent fast sampling
  • Average-to-peak performance extrapolation parameters
  • Correlations with other load indicators (output queue drops, ignores)

Data collection technique

• Based on fast SNMP polling of a production router
  • 1-2 Hz maximum rate (200,000 samples per day)
  • No detrimental router performance impact
• Uses router clock as the rate timer
• A restricted set of variables from only one interface is collected:
  • Router uptime (10 ms time tick interval)
  • Bytes--input and output
  • Unicast packets--input and output
  • Input errors and output queue drops
• Analysis paradigm potentially applicable to other techniques
  • RMON
  • Network Probe Daemon (NPD)

Data acquisition technique caveats

• Router SNMP read-only community string required
• Polling frequency stability important
• Not robust under local network congestion or router CPU overload
• Impact of slow router SNMP counter refresh rates
• SNMP 32-bit counters and OC-48: octet variables can wrap every 14 seconds

Data set parameters

• DMZ FDDI interface of a UW border router
  • Outbound traffic to internetMCI and NorthWestNet
• Thursday, 29 February 1996 0900-1930 PST
  • 50,000 polls
  • Output rate: 3.1 Mbps average, 15.9 Mbps peak
• Time stability
  • Average poll interval: 0.78 +/- 0.08 seconds
  • Polling time distribution
• Data quality measures

Polling time distribution

Data quality measures

Analysis techniques

• Bulk characterization
  • Chronological plot of average and peak rates
    • Peak/average ratio = 1.9 (+/- 15%)
  • Rate distribution
  • Integrated rate spectrum
• Peak characterization
  • Convoluted peak rate versus polling frequency distribution

Chronological plot of average and peak input rates

Measured output rate distribution

Integrated output rate spectrum

Linear plot

Logarithmic plot

Convoluted peak input versus polling frequency distribution

Comments and future directions

• Quick-and-dirty technique geared for ISP/campus network types
• Impact of national backbone bottlenecks and protocol/application modifications
• Connection with self-similar (fractal) analyses of peak loads