ATM NAP Progress - NANOG February 1995

The HTML version provides the content of the slides in a form more convenient for browsing on the web and provides additional detail. In the HTML, the individual slides are separated by horizontal rules. Small inline images of the graphs are provided, but larger images are available by selecting the smaller image (clicking on it for most HTML readers).

• large aggregation of traffic

• mostly low speed (under T1) short lived flows

• high speed applications at times dominate traffic

• Merit (John Scudder) predicts high traffic levels

• NAP performance should not be a step backwards

• Response to congestion is very difficult to quantify

• A goal is to be able to measure or estimate:
  • Aggregate application goodput
  • Fairness

• Link utilization and loss are usefull
  • As a basis for estimates of goodput
  • Zero loss expected at low average utilizations
  • Low loss as high average utilization is approached
  • Loss may become high very near full utilizartion
  • Short term congestion should not cause loss

• Fairness is extremely difficult to measure

• Lowered HSSI clock rate to E1 (34.3 MHz)

• This is a workaround

• TTCP would cause a loss of buffers

• Problem easily reproduced with multiple pings

• Cleared by Power cycle until TTCP was run

• This ADSU problem has been fixed

• Worst case is large followed by small packets

• Example: large MTU and compressed ACKs

• Poor performance with bidirectional congestion

• Little degredation when FDDI MTU is not in use

• Examination of the real traffic mix is needed

• May still be masked by ADSU problems

• Performance has yet to be adequately studied

• Severe fairness problems have been observed

• This is quite speculative

• Adequate performance at current traffic levels

• Large MTU or high speed flows may cause trouble

• Substantial traffic growth may cause trouble

• Fairness problems need study, could be very severe

• The vBNS requirements cannot yet be met

• Measure individual goodputs using tcpdump

• Continue to look at aggregate goodput

• Compare ingress utilizations looking at fairness

• See if SNMP can warn of fairness problems

• AADS has offerred to host a joint testing facility

• Interest in participating has been expressed by:
  • ANS
  • Merit
  • MCI
  • Bellcore

• Goals and preliminary test plans have been written