• ADSUs are troubling.
  • Today
    • Bugs in buffering, clock, segmentation chips.

  • Today & Tomorrow
    • Certain NAP traffic conditions break the ATM VBR model, N to 1 configuration (N > 1, offered load egress > 36.8Mbps).
    • As a result, even a fixed ADSU could perform less than adequately in N to 1 configuration.
    • VBR model requires back pressure in the form of rate shaping
    • Lack of rate shaping in the ADSU makes a large buffer in ATM switch very important.
    • AADS' ATM Switch has 4000 cell buffer (~40-50msec delay at 36.8Mbps)

• Learned about the testing process.
  • Testing is difficult: simulating Internet is impossible, stressing boxes is difficult.
  • Resources are tight (people and hardware) and expensive.
  • Delay on WAN link is very important to simulate.
  • Homogeneous hosts make test results easier to interpret.
  • Just about everything inter-operates: wellfleets to ciscos, cisco DS3 to cisco OC3c, FR wellfleet to ATM cisco, ...

• There are more tests that can be run to better characterize ATM performance.
  • Even ciscos break on windows above 500KB, i.e. the size of the card to card buffer pool in the cisco.
  • Problem is not limited to ATM, ATM just doesn't look great right now because of the smaller buffer.
  • This leads to other tests we can run to better compare ATM performance to known configurations, i.e. SMDS, FDDI.

• What will happen when the ADSU is fixed?
  • We can guess that the congestion collapse will look much better.
  • With fixed DSUs we should be able to stabilize at "near line speed" with before we see congestion collapse.

• How much load will NSPs drive between themselves?
  • Predicting real world conditions are important to arrive at conclusions from test results.
  • Characterizing NAP use figures greatly in an ATM NAP using ADSUs.

• Pool resources.
  • AADS now has lots of hardware.
  • AADS needs more intellectual resources (i.e. Curtis, Bellcore).

• There is a lot of data to correlate when mapping lab performance to the real world.
• Once this is done, we should be able to characterize how the ATM NAPs will perform under various conditions.
• At least enough to hopefully give warm fuzzys...

• Run with "fixed" DSUs/HSSI and gamble?
  • Will multiple NSPs burst for 40ms to same egress and congest (switch breaks)?
  • Will a pair of NSPs burst at each other at a 512K or larger (aggregate) window (cisco breaks)?
  • Will our switch vendor give us an 8000 cell buffer card?

• Co-location option?
  • Heavy financial overhead.
  • Co-location is not easy.
  • Co-location is not viable in the near term.

• Use AIPs?
  • DS3 AIPs are vapor but should be a good solution if DS3 vs. OC3c pricing is deciding factor.
  • OC3c AIPs are ideal:
    • They rate shape.
    • Lots of bandwidth with which to work.
    • Good SCR between NSPs.
    • They work today (?).

Stan Barber's Notes

Pooling resources will help make the future of the ATM NAPS brighter. AADS will provide a place for everyone to get into one place and test. Most of this was already discussed by Mark earlier. The important thing to figure out is what configurations will work. Once AADS gets the new ADSUs, they will go back in the lab and test. AADS believes that OC3c will work just fine today if the Cisco AIP works just fine.