This panel exists because KC complained at the last NANOG that there seemed to be no resolution on where loss is happening. This forum will bring together providers, NAP operators and R&E folks to see if there can be some way to resolve this.
Steve Feldman of WorldCom (formerly MFS) presents next.
There is packet loss at the MAEs. Head-of-Line Blocking, Primarily on inter-switch trunks, difficult to pin down because traffic patterns change.
In a crossbar switch, input queues are sending packets to output queues. If there is too much traffic heading to one output queue then traffic backs up waiting for the target output queue to drain.
How can this be measured? Using SNMP, you can measure ifInDiscards, comparing ifInOctects and ifOutOctets and doing ping tests.
Short-term
Long-Term
Peter Lothberg notes that long ago MCI, Sprint and UUNET approached MFS at MAE-EAST and they cut a deal for a $1000 one-time fee for private interconnects at the MAE. Steve notes that this never made it into any official company policy, but he is working to get a reasonable policy in place to permit private interconnects at the MAEs.
Steve notes that the future will need to accommodate OC-12 ATM access. He may have more to say about this next NANOG.
Someone asks, "What is the plan for expanding the capacity between AMES and MFS at MAE-WEST?" There are three OC-3s now and one more is aviailable when traffic indicates a need it will be put on-line.
KC is back. She notes that both the backbone providers and the NAP operators are both probably correct in saying that there is no lose in the backbone and no lose in the NAPs. It appears clear that the problem really is in the interconnection between the two.
Warren Williams notes that there are a number of issues in controlling this. Many of the large companies are not able to get adequate provisioning into some NAPs because they have such large needs for provisioning, that they may not be able to get adequate provisioning into some NAPs. Some of the small companies are more aggressive in addressing these needs.
Steve Feldman notes that at the MAEs, the load is beginning to stretch the limits of a FDDI-based infrastructure.
Steve Schnell says that if you use a full-duplex FDDI card, you can get more throughput.
KC notes that having a traffic matrix in the GIGAswitch seems to be a critical mechanism to help facilitate optimization of the GIGAswitch-based NAPs. Stephen Polit from DEC will be speaking to this and other issues involving the GIGAswitch later today.
Steve notes that this is not a complete solution, but it will not force those customers will inadequate link sizes to upgrade.
What about traffic matrix information from ATM? Warren says that this is possible. There are issues of disclosure that have to be addressed (Worldcom has the same issues at the MAEs).
What about co-location space upgrades at the MAEs? Steve does not know the answers for this. This is a responsibility of the transmission people.
Both PAC*BELL and WorldCom noted that some things have to be referred to sales. However, most contracts from permit loss. Is there something wrong with the contracts? Steve says that from an engineering perspective he would not want to run a MAE that routinely drops packets and from an business perspective, a badly performing MAE will cause customers to migrate away.
Warren Williams notes that peering agreements and transit agreements are not part of the relationship between the ISP and the NAP operators. This makes is hard for packet-loss liability to be clearly assigned. However, that also means that if access-lines are the problems, then the language in the peering and transit agreements needs to deal with this.
Curtis asks if ANS could get access to stats on his PVCs from PAC*BELL. Warren says yes and to talk directly to him about it.
KC notes that it looks to the government like the folks running the Internet don't have their stuff together. That may result in regulation.