The IP Over NBMA (ION) Archive

[Masataka Ohta <mohta@necom830.hpcl.titech.ac.jp>]

Jeremy;

> We're nearing the point where
> announcing entirely new proposals is counter-productive.

What is counter productive is to propose or develop inscalable
protocols for CSRs such as IFMP, TAG switching, SITA, ARIS...

CSRs with resource reserved traffic need NO new protocol.

Just use RSVP (with LIH field) and Q.2931 (with BHLI field) and
everything, including multicast, works just fine in a way independent
to any existing routing protocol.

As for best effort traffic, the only known way so far of scalabily
switching the best effort traffic is to use "step", which I
architected at the same time I architected the CSR in early 1994
and made the idea public about a year ago in some ML discussion.

What is "step"? See an attached extended abstract sent for INET'97.

Discussion on "step" is welcome in the existing mailing list on
CSRs:

	colip-atm@necom830.hpcl.titech.ac.jp

						Masataka Ohta
---
Hop, Step, but don't Jump for Scalable Lower Layer Forwarding
of Best Effort Traffic

1. Introduction

A scalable approach to reduce the load of layer 3 header processing
by stepping through certain number of intermediate routers is
proposed.

Lower layer forwarding is a technique to forward a lower layer unit
of data over the Internet looking only at the layer 1 or layer 2
label to reduce the load of layer 3 header processing [CSR1, CSR2,
IPS, TAG]. To do so, it is necessary to maintain a correspondence
between layer 3 final or intermediate destination and the lower
layer label, so that the lower layer label can uniquely identify
the layer 3 destination.

For example, on CSRs (Cell Switching Routers) [CSR1, CSR2], ATM
VPI/VCI is the lower layer label to uniquely identify an RSVP/ST2
flow. So, it is possible to use RSVP/ST2 to setup cell switching
fabric on a CSR. Note that, while the terminology "layer 2 forwarding"
is often used, ATM VPI/VCI changes physical layer by physical layer
and is a layer 1 label, which is why this paper use the terminology
"lower layer forwarding". Note also that, with IP switching [IPS]
where a layer 2 segment always consists of a single layer 1 segment,
ATM VPI/VCI is also a layer 2 label.

2. Why not to jump for best effort traffic

With resource reserved communication, "jump", that is, establishing
lower layer forwarding all along the path between the source and
the destination is worth doing [CSR1, CSR2].

But, it is not so a good idea for best effort traffic. While there
are several proposals [CSR2, IPS, TAG] to use lower layer forwarding
for best effort traffic by automatically detecting a flow, or a
frequently used pair of source and destination, there is a scalability
limitation in such approaches.

That is, as the size of the network grows, the number of flows also
grows to be larger than the number supported by the lower layers.

For QoSed flow, it is not a problem because it is expected that
certain amount of bandwidth is also reserved, which assures that total
bandwidth limitation will be a severer restriction. For example, over
a link of 156Mbps, at most 2437 64Kbps reservations can be made.

But, with best effort traffic, there may be unlimited number of flows
exist. Moreover, as the load to the network increases, the speed of
each link is lowered and the number of flows increases. Considering
that the Internet today have about 10**5 toplevel routing table entry,
it is unlikely that a valuely suggested aggregation of flows in [TAG]
works so well.

Of course, we can always give up lower layer forwarding and fall back
to normal but less efficient hop-by-hop forwarding. But, it is a bad
behavior that the performance degrades at the time of congestion where
the performance is most needed.

For lower layer forwarding meaningful, we should forward as many
packets as possible, at the lower layer, which means to support
a lot of flows.

3. Stepping through several routers

A "step", here, means a lower layer channel directly connecting
two routers bypassing several, but not so many, routers.

With certain routing protocol such as OSPF, it is possible to know not
only the next but also the second, the third or even the forth next
hop router. Thus, by establishing "steps" from the current to the N-th
next hop router, it is possible to reduce the average number of IP
header processing by 1/N.

If a router has D neighbor routers, D*(D-1)**(N-1) channels are
necessary to have N-hop steps (N>=2). For example, to support upto
5-hop steps with 4 neighbor routers, to reduce the IP header
processing load by 80%, only 480 channels are necessary.

Note that the mechanism needs no detection of flows.

Moreover, lower layer forwarding depends only on the local topology
and stable regardless of the size of the entire network.

To reduce the load of router at the border of lower level routing
domains of hierarchical routing or of firewall routers, it is
necessary to make the border fuzzy to distribute it over several
routers (firewalls).

References:

[CSR1] M. OHTA, H. ESAKI, K. NAGAMI: "Conventional IP over ATM",
   <draft-ohta-ip-over-atm-00.txt>, Internet Draft, March 1994.
[CSR2] H. ESAKI, K. NAGAMI, M. OHTA: "High Speed Datagram Delivery
   over Internet using ATM Technology", Networld+Interop '95 Engineer
   Conference,  E12-1~E12-9, 1995.
[IPS] P. W. Edwards, R. E. Hoffman, F. Liaw, T. Lyon, G. Minshall,
   "Ipsilon Flow Management Protocol Specification for IPv4  Version  
   1.0", RFC 1953, May 1996.
[TAG] Y. Rekhter, B. Davie, D. Katz, E. Rosen, G. Swallow, "Tag
   Switching Architecture Overview", <draft-rfced-info-rekhter-00.txt>,
   Internet Draft, September 1996.