NANOG Panel Discussion: Are

1
NANOG Panel DiscussionAre Transit
Exchangesand Peering ExchangesSelf-Differenti
ating?

• Toronto
• June, 2002

2
Keith MitchellChief Technical OfficerXchangePoin
tPhil SmithRouting SpokesmodelCiscoBill
WoodcockResearch DirectorPacket Clearing House
3

• Background white-paper for this discussion
• www.pch.net/documents/papers/transit-exch-peering-
  exch

4
Past Assumptions

• All exchanges are functionally similar, differing
  principally qualitatively.
• People go to exchanges to peer, but buy transit
  at their own location or within a colocation
  facility.

5
Observed Reality

• Some exchanges are vastly more expensive than
  others, and provide services which are
  economically unsupportable within a peering
  economy.
• Many parties buy and sell transit within exchange
  facilities, either over crossconnects or through
  the switch fabric.

6
Hypothesis

• There are actually two poles toward which
  exchanges can be optimized peering or transit,
  and they pose very different economic and
  provisioning requirements.
• It may not be possible to meet both requirements
  with maximal success within the context of a
  single exchange.

7
Aggregation Benefits

• An aggregation benefit is an efficiency derived
  from grouping traffic from many sources or
  destinations together into one shared connection,
  rather than putting each divisible unit of
  traffic through its own connection.
• Aggregation benefits typically result in
  increased apparent performance or cost savings or
  both.

8
Aggregation Benefits (2)

• The down-side of aggregation is that it
  aggregates not only the traffic, but it
  simultaneously aggregates risk.
• There is thus a tradeoff between the
  irreplaceability of a circuit and the degree of
  aggregation benefit which can be recouped from
  it.
• In other words, a path which carries critical
  routes must have redundancy.

9
Exchange Points and Aggregated Risk

• Peering is an economic optimization. That is,
  peering provides a lower-cost route to a
  destination otherwise reachable through transit.
• A route learned through peering is always
  accompanied by a redundant transit path, whereas
  a route learned through transit may not also be
  reachable through a peering path.

10
Exchange Points and Aggregated Risk (2)

• Further redundancy is never necessary on a
  peering path. Full aggregation benefit may be
  extracted from it without danger, since the
  redundant transit path provides a safety net.
• Transit contains full routes, including those not
  learned through peering, and constitutes a sole
  means of reaching most destinations. It thus
  requires redundancy and reliability. Aggregation
  benefits cannot be fully realized on a transit
  link.

11
Peering Exchange
Transit Exchange

• Maximization of aggregation benefits suggest
  that
• exactly one peering exchange exist in a region,
  and
• that it include all available peers.

• Reliability by dint of redundancy suggests that
• two or more transit exchanges exist in a region,
  and
• that each can operate competitively with no more
  than three buyers and three sellers.

12
Peering Exchange
Transit Exchange

• Making all peers available at one exchange
  dictates that the exchange switch fabric within a
  region never be fragmented, regardless of how
  many exchange point operators exist.

• Since at least two transit exchanges should
  exist in a region, different operators switch
  fabrics need not be interconnected.
• This allows transit exchange operators to more
  closely govern the quality of their service.

13
Peering Exchange
Transit Exchange

• Aggregation suggests that peers should be
  reached through the switch fabric, up to the
  limit of the capacity of the fastest available
  switch port.

• Reliability suggests that if only a couple of
  transit relationships are needed within an
  exchange, direct interconnections are the best
  means of facilitating them.

14
Peering Exchange
Transit Exchange

• Maximum benefit is reached at the lowest price.
• The price is bounded by the additional cost of
  accommodating otherwise-peered traffic through
  transit during peering exchange downtime.

• Maximum benefit is reached at the highest
  reliability.
• The price arbitrages the difference between the
  sum of the market sale price plus backhaul and
  the cost of transit delivered at the customer
  premises.

15
This Begs the Question

• If a provider is already at two exchanges, to
  purchase transit in a redundant fashion
• and the marginal cost of pushing additional
  traffic through those two exchanges is less than
  that of going to a third exchange, even a very
  cheap one, then
• have we just gone full circle? Will peering
  become an add-on service of transit exchanges?