Internet Draft

Internet-Draft Arup Acharya, C&C Research Labs, NEC USA
Expiration Date: Rajiv Dighe, C&C Research Labs, NEC USA
January 1998 Furquan Ansari, C&C Research Labs, NEC USA
July 1997

IPSOFACTO: IP Switching Over Fast ATM Cell Transport

<draft-acharya-ipsw-fast-cell-00.txt>

Status of This Memo

This document is an Internet-Draft. Internet-Drafts are working
documents of the Internet Engineering Task Force (IETF), its areas,
and its working groups. Note that other groups may also distribute
working documents as Internet-Drafts.

Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."

To learn the current status of any Internet-Draft, please check the
"1id-abstracts.txt" listing contained in the Internet-Drafts Shadow
Directories on ftp.is.co.za (Africa), ftp.nordu.net (Europe),
munnari.oz.au (Pacific Rim), ds.internic.net (US East Coast), or
ftp.isi.edu (US West Coast).

Abstract

This document describes a method for mapping IP flows to ATM
switches. No signaling is necessary to setup a path through ATM
switches. Switch controllers run a IP routing protocol and execute IP
forwarding. The Ipsofacto component is responsible for mapping a IP flow
to a switched path. The focus of this document is primarily on switching
IP multicast. Mechanisms for switching unicast flows are also described.

Table of Contents
1. Introduction
2. Basic operation
3. Related Schemes
4. Switch configuration
5. Protocol Independent Multicast-Dense Mode
6. Protocol Independent Multicast-Sparse Mode
7. Switching Unicast flows
7.1 TCP Optimizations
7.2 Fixed Length IP routing tag
7.3 Optimizing Mobile-IP flows
7.4 Aggregation through IP-in-IP encapsulation
7.5 Merging aggregated flows through a
multipoint-to-point tunnel
7.6 Handling Route Changes

Acharya, Dighe, Ansari Expires: January 1998 [Page 1]

8. Conclusions
9. Security Considerations
10. Intellectual Property Considerations
11. Acknowledgments
12. References
13. Authors' Addresses

1. Introduction

IPSOFACTO is a methodology for executing the IP protocol stack
directly on ATM switches. The ATM signaling stack is not used. This
draft describes how IP multicast and unicast may be mapped directly to
ATM switches.

The target scenario consists of (a) a cloud of ATM switches running
the IP protocol stack and optionally, (b) hosts that are directly
connected to this cloud. Each switch in this cloud runs the IP
protocol stack along with an Ipsofacto component for mapping IP flows
to switched paths. The description in this draft assumes
configuration (a).

2. Basic Operation

The basic premise of operation for Ipsofacto is that all unused VCs on a
input port of a switch are mapped to the switch control processor
(Fig. 1).

---------------------
| IP | IPSOFACTO |
---------------------
| Switch Controller |
---------------------
|
| \
/|\ \
/ | \ \_______unused VC=1
Port1 --/-----\--- Port3
----------- | /Switch \ |------------
-/---------\-
__________/ | \
unused VC=1 | |
Port2 | unused VC=1

FIGURE 1

A cell-level switched path for data forwarding is established in the
following way. A sender selects an unused VC on an outgoing link to
forward the first packet of a new flow. This is received by the switch
processor at the downstream end of the link, which then selects

Acharya, Dighe, Ansari Expires: January 1998 [Page 2]

outgoing links based on its IP routing tables, and an unused VC for
each selected link. This first packet is then forwarded by the
processor on the selected outgoing links by picking an unused VC on
each link (Fig. 2).

---------------------
| IP | IPSOFACTO |
---------------------
| Switch Controller |
---------------------
========> (IP forwarding)
Pkt1 | Pkt1
/ | \
/ | \.............unused VC=1
Port1 ---/-------- Port3
-----------| / SWITCH |------------
------ -/---------
| Pkt1.........../ |
----- unused VC=1 |
Port2|

FIGURE 2

Subsequently, the switch processor adds an entry in the VC tables,
<{input port, input VC}, {output port, output VC}>. Following packets
are then switched at the cell level, eliminating the need for
packet-level forwarding at the control processor.

---------------------
| IP | IPSOFACTO |
--------- ||---------
| Switch Controller |
--------- || --------
| ||
| || ADD switched path
| || (port1, VC=1 --> port2, VC=1)
| ||
|----- ||----|
| __\/__|________
-----------| / SWITCH |------------
Port1 |--/---------| Port3
___________________/ |
|
Port2|

FIGURE 3

In contrast to data packets, which are forwarded through a switched
path (following the first packet), a switched path is never created

Acharya, Dighe, Ansari Expires: January 1998 [Page 3]

for IP "control" messages: typically, IP "control" messages will be
sent and received on a predefined "control VC", and will therefore, be
forwarded through the switch processor. As a result of processing
these control messages, a per-flow forwarding state may be established
on the switch processor. Changes in the forwarding state,e.g. pruning
an outgoing interface, is then used to modify the switched path
(e.g. deleting an from the VC tables). When the
forwarding state is removed from the control processor, the
corresponding switched data path is released, by marking the input and
output VCs as unused. For PIM, the "control" packets are distinguished
from data packets, based on protocol number in the IP header. For TCP,
the "control" packets may be decided based on flags in the TCP header
(SYN/FIN). ICMP packets are always forwarded through the control
processor.

When an outgoing VC, that is currently in use, is reclaimed back,
i.e. marked as 'unused', this action is preceded by marking the VC to
be in 'transient' state, sending a RECLAIM message from the upstream
to the downstream node, and waiting for a RECLAIM-ACK in reply. When
the downstream node receives a RECLAIM message, it marks the
corresponding incoming VC as 'unused', responds with a RECLAIM-ACK,
followed by sending RECLAIM messages further downstream. The RECLAIM
message may need to be re-sent if a RECLAIM-ACK is not received within
a specified interval. Once the RECLAIM-ACK is received, the VC is
marked as 'unused' at the upstream node. Note that it may not be
necessary to send a per-VC RECLAIM and RECLAIM-ACK, and instead, a
single RECLAIM and RECLAIM-ACK message may contain multiple VC
numbers; this cuts down the inter-switch control traffic at the
expense of delayed reclamation.

A RECLAIM message is triggered as a result of some protocol action at
the controller, e.g. an outgoing branch of a switched path may be
removed because a PIM "Prune" message was received.

It is not necessary that the RECLAIM is only sent from the upstream to
the downstream node; the downstream node may initiate the RECLAIM as
well.

In this draft, the terms 'marking a VC unused' or 'reclaiming a VC'
implicitly imply that RECLAIM and RECLAIM-ACK message exchange has
been completed.

3. Related Schemes

Unlike MPOA [MPOA], Ipsofacto does not use ATM signaling to setup
switched paths for IP flows. Ipsofacto is similar in spirit to Ipsilon's
IP switching [Ipsilon] [IFMP] and Toshiba's CSR [CSR]. Ipsilon selects
the VC for switching a flow from the downstream node, while CSR
selects the VC at the upstream node. Both use a notification protocol
(IFMP for Ipsilon, FANP [FANP] for CSR) to inform the upstream/
downstream neighbour of the selected VC. Ipsofacto informs the
downstream node of the selected VC implicitly through usage. Other
related proposals include Tag-switching [TAG] and ARIS [ARIS]: both
pre-establish switched paths based on routing information.

Acharya, Dighe, Ansari Expires: January 1998 [Page 4]

4. Switch configuration

Each port of a switch is assigned to be an IP interface. Incoming
cells on all unused VCs are directed to the control processor. When a
packet is received on a unused VC, the incoming VC and port number
information associated with the packet is made available to the
Ipsofacto layer. Ipsofacto layer keeps track of all unused VCs on each
port, and maintains a flow-table which maps a flow to a switched path.
e.g.
---> <{input port, input VC}, list of {output port, out VC} pairs>
The flow-id is constructed from the source/destination IP addresses
and port numbers along with the incoming TTL value on the packet.

Although not strictly necessary, we will assume that there is a
dedicated control VC between neighbouring switches. Typically, the
control VC will be used to forward IP multicast "control" messages
(i.e. PIM packets with protocol number 103) between adjacent switches.

In effect, each switch together with its IP/Ipsofacto layers appears
as a network router to its adjacent switches/hosts.

5. Protocol Operation: Mapping PIM Dense Mode [PIM-DM] flows in Ipsofacto

When there exists no forwarding cache entry at a switch controller for
a multicast flow, the first data packet for that flow will install a
forwarding cache for the flow at the controller, as specified by the
protocol operation for Protocol-Independent-Multicast(Dense Mode).
The VC# and the port number (selected by the upstream switch
controller) on which the data packet is received, is available to
Ipsofacto. Additionally, for each outgoing interface in the newly
created multicast forward cache, Ipsofacto selects an unused VC to
forward the packet to the downstream switch controller. At this time,
it adds an entry in the flow-table, mapping the flow-id to a switched path {, list of }, and adds entries in the
hardware VC routing tables corresponding to {,
list of }. Subsequent packets in the flow are
not processed by the controller, and are switched at the cell level
making use of the hardware multicast capability of the ATM switching
fabric.

In essence, the IP multicast routing protocol is used to establish the
multicast forwarding state at the switch controller,while Ipsofacto
maps this state to a point-to-multipoint VC within the switching
fabric.

In PIM-DM, the forwarding state for a flow is associated with an
expiration timer. This timer is restarted whenever a data packet is
being forwarded. However, in Ipsofacto, packets subsequent to the
first are not forwarded through the switch controller. Therefore, the
rule for handling a timer expiration is as follows: when the timer
expires, the controller checks the corresponding switched path for any
cell forwarding activity since the last expiration. If cells have been

Acharya, Dighe, Ansari Expires: January 1998 [Page 5]

forwarded in the interim period, the timer is restarted. Else, both
the forwarding state and the switched path are deleted.
The list of outgoing interfaces associated with a forwarding cache
entry, may be modified by arrival of Prune and Graft messages. The
corresponding switched path is modified as follows: for a Prune
message, the for the interface on which the
Prune is received, is deleted from the VC tables, i.e. the
corresponding branch of the point-to-multipoint VC is deleted. If
after deletion of the branch, no more outgoing branches remain
(i.e. the forwarding state is now a negative cache entry), then the
switched path is deleted, and the the incoming VC is marked as unused,
i.e. it points to the switch controller. When a Graft message for a
group G is received, PIM-DM will add the received interface to the
outgoing interface list for all (S,G) forwarding entries. Ipsofacto
complements this action, by picking an unused VC and adding a branch
to the corresponding switched path, for each
of the (S,G) entries modified by the Graft message. If the (S,G)
entry is a negative cache entry, then no switched path will exists for
the (S,G) entry and Ipsofacto will not take any action; a new switched
path will be created when with the arrival of the first packet.

Note that Ipsofacto does not change PIM protocol actions associated
with PIM control messages; Ipsofacto only determines how the resulting
multicast forwarding state needs to be mapped to a switched path.

Ipsofacto supports Distance-vector multicast routing protocol (DVMRP)
[DVMRP] in a similar fashion.

6. Mapping PIM Sparse Mode [PIM-SM] operation under Ipsofacto

PIM-SM creates a shared forwarding tree rooted at a Rendezvous Point
(RP) with receivers at the leaves. It allows for individual receivers
to bypass the shared tree and instead receive data packets on a
source-specific tree. Unlike PIM-DM, forwarding entries at
intermediate routers are created by sending explicit Join messages
towards the RP (shared tree) or the source (source specific tree),
from last-hop routers.

Ipsofacto maps the shared forwarding entry on the RP-tree to
per-source switched paths at intermediate switches. This is a slight
departure from traditional packet-level routers which maintain a
common (*,G) packet-level multicast forwarding state: under Ipsofacto,
the forwarding state at the switch processor is shared by all (*,G)
flows, but cell-level switched paths are kept separate for each (S,G)
flow. The reason for separating the switched paths is that PIM-SM uses
the shared tree to reach all receivers by default, while
simultaneously using part of the shared tree to reach a subset of the
receivers for specific senders. Therefore, an intermediate node on the
tree may contain forwarding entries of the form (S,G) (with RPT-bit
set) and (*,G) : though the incoming interface for both entries may be
the same, their outgoing interfaces will differ. The decision whether

Acharya, Dighe, Ansari Expires: January 1998 [Page 6]

to use a (S,G) entry or (*,G) entry is taken on a per-packet basis,
using the source address of the packet. But, such a decision cannot be
made when forwarding data (cells) in a switched path, since the data
packet is not forwarded through the switch processor . Hence, under
Ipsofacto, instead of using a common point-to-multipoint switched
path, a separate switched path is used for each multicast flow.

Under Ipsofacto, the first data packet for a multicast flow is handled
similarly for both PIM-DM and PIM-SM. The outgoing interfaces, on
which to forward the packet further, is selected from the multicast
forwarding entry, and a switched path is added to the VC tables. A
corresponding entry (, list of ) is added to the flow-table. Unlike PIM-DM, where each
multicast forwarding entry corresponds to a single entry in the flow
table, each multicast forwarding entry (e.g. (*,G)) in PIM-SM may be
associated with multiple flows (switched paths) and therefore,
multiple entries in the flow-table. This one-to-many association
between multicast forwarding entries and switched paths is explicitly
recorded in the flow-table. When a Prune or a Join message affects a
multicast forwarding entry, the associated entries in the flow-table
will be modified accordingly, e.g. if an outgoing interface is pruned,
then for each switched path associated with that forwarding entry, the
corresponding will be deleted from the switched
path. Analogously, for a Join, an unused VC will be selected on the
outgoing port (interface) and added to the switched paths associated
with the affected forwarding entry.

There are two special interactions that need to be considered for
PIM-SM. The first occurs when an intermediate router is in the midst
of switching from the shared tree to a source specific tree for a
particular (S,G) flow. When the first data packet arrives on the
source-specific tree, PIM-SM triggers a Prune message to be sent on
the shared tree. Under Ipsofacto, this implies that the existing
flow-specific switched path associated with the shared tree will no
longer be used: the switched path (and the flow table entry) may
directly be reclaimed when the Prune message is sent, or allowed to
time out. Note that the arrival of the data packet on the
source-specific tree will lead to a new switched path being setup.

The second interaction occurs at the Rendezvous Point (RP). The RP may
receive data packets from a sender either encapsulated as unicast
(Register) packets or as native multicast (which typically happens
when the RP has joined the source-specific tree). In the former case,
the RP cannot created switched paths for specific flows: it must
decapsulate the packet first. Further, the flow table entry will now
contain a list of pairs; there is no incoming
port,VC. For packets belonging to the same flow, the RP must use the
same VC on a outgoing port for successive packets. This VC is recorded
in the flow table. In the latter case, i.e. when the RP receives a
native multicast packets, it creates a switched path as mentioned
earlier.

Acharya, Dighe, Ansari Expires: January 1998 [Page 7]

7. Switching unicast flows

Mapping the first data packet of a unicast flow to a switched path is
similar to that for a multicast flow. IP routing determines the
outgoing port, Ipsofacto selects an unused VC on that port and adds a
switched path {, }.
Unlike the case for multicast, where PIM control messages are
implicitly used to refresh the switched path, switched paths for
unicast flows require a separate refresh protocol.

The downstream endpoint of a switched path, periodically monitors
traffic on the , checks if the received packet header
matches the expected header, and if so, sends a REFRESH message to the
upstream node. Intermediate nodes need only forward the REFRESH
message upstream; it is not necessary to locally monitor traffic
activity. When the downstream endpoint of a switched path decides not
to send a REFRESH message upstream, it may optionally send a RECLAIM
instead to speed up the process of path teardown.

Other alternatives to refresh the switched paths are also possible,
such as using IFMP or FANP for refreshing the switched unicast paths.

7.1 Optimizations for TCP flows

For TCP flows, data transfer is preceded by a three-way handshake.
Under Ipsofacto, the switched path is setup when the first packet
(SYN) is forwarded; thus, an end-to-end switched path is available
prior to data transfer. The teardown of the switched path (i.e marking
the switched path VC as unused at each hop) can be hastened by sending
the FIN and the ack to the FIN packets through the switch controller
instead of the switched path, i.e these packets are treated as
"control" packets. The switched path can be reclaimed at a node after
the FIN from each endpoint, and the corresponding acks have been
forwarded through the controller. This optimization works only at
nodes where the forwarding path for the TCP flow is symmetric,
i.e. the incoming port for one direction of the flow is same as the
outgoing port for the reverse direction, and vice-versa.

7.2 FLIRT (Fixed Length Ip Routing Tag)

The first packet of a flow requires an IP routing table lookup at each
hop within the Ipsofacto cloud. This longest-prefix match can be
replaced by an exact match on fixed number of bits by attaching a
fixed-length tag with each packet. The tag-switching architecture
[TAG] uses a tag distribution protocol [TDP] to distribute a
consistent set of tags. However, when using ATM switches as
Tag-switch-routers (TSR), the tag also defines the VC on which the
packet is forwarded under the Tag Switching architecture.

Under Ipsofacto, the VC selected for forwarding a packet is decoupled
from the tag: the VC selection is implicitly by usage, as explained
earlier. Regardless of the VC selected on the outgoing link, the
upstream switch may optionally insert a FLIRT before the IP packet

Acharya, Dighe, Ansari Expires: January 1998 [Page 8]

within a AAL5 frame. (Absence of FLIRT is represented by zero-ing out
the bits of the fixed length field). The FLIRT on an incoming packet
is then used to decide the outgoing FLIRT and the outgoing port
without doing a IP table lookup. Before forwarding the packet, the
FLIRT is replaced with the new value.

If the FLIRT is absent (as may be the case when the first packet is
lost), normal IP route lookup is performed.

In future, the FLIRT may be extended to hold additional fields,
e.g. id of the ingress point to the Ipsofacto cloud, flow type (host
pair, port pair, IP-in-IP), QoS parameters.

7.3 Optimizing Mobile-IP flows

In scenarios where a switch processor acts as the home agent for a
mobile host[Mobile-IP], it may optionally speed up the process of
tearing down a switched path (leading to the previous location of the
mobile host). For flows initiated from a correspondent host, the
packet is addressed to the mobile host's home address, which is then
intercepted by the home agent when the mobile host is not at home.
The home agent then forwards the packet after encapsulating with the
mobile host's current foreign address. Under Ipsofacto, this leads to
two switched paths, one from the correspondent host to the home agent
and the second, from the home agent to the mobile's current location.
The second switched path is setup by using the outer
source/destination address of the encapsulated packet as the flow-id
(i.e there are no transport port numbers), and in effect, aggregates
all flows to the mobile into a single switched path.

When the mobile host changes its foreign address, it must send a
Registration Notification to its home agent; the home agent replies
with Registration Reply to the mobile host's foreign address. If the
registration is accepted, then the switched path leading to the mobile
host's previous foreign address is no longer useful. In this
situation, Ipsofacto may optionally send RECLAIM messages to the
downstream nodes of the existing switched paths leading to the mobile
host's previous location. The new switched path is automatically setup
when the first packet addressed to the mobile host's new foreign
address is forwarded.

For each incoming flow, in general, it is not possible to bridge to
the incoming switched paths and outgoing (aggregated) switched path,
since the home agent must encapsulate each packet with the mobile's
foreign address. Specific optimizations are presently under study.

7.4 Aggregating multiple unicast flows into a single switched path

As the previous section points out, it is possible to use IP-in-IP
mechanism to aggregate multiple flows with a common pair of entry and
exit points into the Ipsofacto cloud. This requires support from the
routing protocol operating in the Ipsofacto cloud, for the entry point

Acharya, Dighe, Ansari Expires: January 1998 [Page 9]

to learn the exit point for a given flow; all flows with a common exit
point can then be aggregated using IP-in-IP encapsulation.

It is also possible to use VP switching within the Ipsofacto cloud,
instead of using IP-in-IP encapsulation. In this case, the entry
router would assign a different VC# for a new flow, but assign a VP#
already in-use for flows that share a common exit point. The VPI field
is used to switch flows with a common entry and exit routers, within
the cloud, while the VCI field is used to locally demultiplex
individual flows at the exit router. This would require the exit
router to "learn" and cache the flow-to-VC mapping: when a new flow is
received at the exit router for which it has no flow-to-VC binding, it
installs such a binding from the IP header and the incoming VC. For
each VC that is bound to a flow, it must then periodically reassemble
cells from that VC to a packet and refresh/replace the corresponding
flow-to-VC binding. Alternatively, a specific VC may be reserved
within each VP for the entry point to periodically update/refresh the
exit point of current flow-to-VC bindings.

7.5 Merging encapsulated flows: multipoint-to-point aggregation

The previous sub-section briefly outlined how multiple flows with
common entry and exit points may be aggregated to a single flow using
IP-in-IP encapsulation. Multiple such flows, each from a different
entry point, but with a common exit point out of the Ipsofacto cloud,
may also be merged in the following way: at a switch N, which has an
existing switched path for an IP-in-IP flow F to exit point B, may
select the same outgoing VC as F for a new flow F1 to the same exit
point B. First, this requires VC-merge capability at intermediate
switches to avoid cell-interleaving. Second, when the REFRESH message
from B is forwarded on the reverse path (as that of the switched
path), a hop-count needs to be included in the message, which is
incremented after each hop. This enables B to decide if the remaining
TTL value on the new incoming flow is no less than the hop-count on
the REFRESH message. If so, the incoming flow may be merged into the
existing flow. Third, at each merge point, such as N, traffic activity
on each incoming branch needs to be monitored periodically. When a
REFRESH message is received from the downstream node, N forwards it
along each incoming branch after incrementing the hop count if there
was traffic activity on that branch within the last k periods.

7.6 Routing table changes/ Loop detection and/or avoidance

Solutions for efficiently handling changes in IP routing tables
require further study. One approach is to leave an existing switched
path untouched in spite of a change in the corresponding IP route, if
intermediate switch controllers on the path continue to receive
REFRESH messages from downstream. Alternatively, when a routing table
entry is updated, all affected switched paths may be forced to go
through IP forwarding by marking the incoming and outgoing VC(s) of
the path as unused.

Acharya, Dighe, Ansari Expires: January 1998 [Page 10]

8. Conclusion

This draft presented a scheme for mapping IP flows to ATM switches,
with special emphasis on IP multicast. Mechanisms for mapping unicast
flows, refreshing unicast switched paths, handling flows leading to
mobile hosts and fixed length tags to speed up routing were outlined.

9. Security Considerations

Security considerations are not addressed in this document.

10. Intellectual Property Considerations

NEC may seek patent or other intellectual property protection for some
aspects discussed in this document.

11. Acknowledgments
The authors would like to acknowledge the following people for
discussions on the contents and/or inputs to this draft: Dipankar
Raychaudhuri, Kojiro Watanabe, Bala Rajagopalan, Kazuhiko Isoyama,
Toshiya Aramaki, Ajay Bakre, Atsushi Iwata and Hiroshi Suzuki. The
authors thank Hideyuki Sakaguchi, Joe Darabpour, Bun Mizuhara, and
Youichi Ohteru for supplying us with switch software for prototyping
Ipsofacto.

12. References

[DVMRP] T. Pusateri, "Distance Vector Multicast Routing Protocol",
Internet Draft , Juniper Networks,
February 1997

[TAG] Y. Rekhter, B. Davie, D. Katz, E.Rosen, G. Swallow, "Cisco
Systems' Tag Swicthing Architecture Overview", RFC 2105, Cisco
Systems, Inc., February 1997

[TDP] P. Doolan, B. Davie, D. Katz, Y. Rekhter, E. Rosen, "Tag
Distribution Protocol", <draft-doolan-tdp-spec-01.txt>, Cisco
Systems,Inc., May 1997

[FANP] K. Nagami, Y. Katsube, Y. Shobatake, A. Mogi, S. Matsuzawa,
T. Jinmei, H. Esaki, "Toshiba's Flow Attribute Notification Protocol
(FANP) Specification", RFC 2129, Toshiba R&D Center, April 1997

[CSR] Y. Katsube, K. Nagami, H. Esaki, "Toshiba's Router Architecture
Extensions for ATM : Overview" , RFC 2098, Toshiba R&D Center,
February 1997

[ARIS] Arun Viswanathan, Nancy Feldman, Rick Boivie, Rich Woundy,
"ARIS: Aggregate Route-Based IP Switching", Internet-Draft,
<draft-viswanathan-aris-overview-00.txt>, IBM Corp., March 1997

Acharya, Dighe, Ansari Expires: January 1998 [Page 11]

[IFMP] P. Newman, B. Edwards, R. Hinden, E. Hoffman, F. Ching Liaw,
T. Lyon, G. Minshall, "Ipsilon Flow Management Protocol Specification
of IPv4 Version 1.0", RFC 1953, Ipsilon,Sprint, May 1996

[Ipsilon] P.Newman, G. Minshall and T. Lyon, "IP switching: ATM Under
IP", Submitted to IEEE/ACM Transactions on Networking

[PIM-SM] D.Estrin et al. "Protocol Independent Multicast-Sparse Mode
(PIM-SM): Protocol Specification", RFC 2117

[PIM-DM] Steven Deering et. al. "Protocol Independent Multicast
version2, Dense Mode Specification", ,
IETF Draft, May 1997

[MPOA] ATM Forum, "Multiprotocol Over ATM version 1.0 (Baseline Text
Version 16)", Andre N. Fredette, Editor.

[IPIP] C. Perkins, "IP Encapsulation within IP", RFC 2003, October
1996

[Mobile-IP] C. Perkins, Editor, "IP Mobility Support", RFC 2002,
October 1996

13. Authors' Addresses

Arup Acharya
C&C Research Labs, NEC USA
4 Independence Way, Princeton
NJ 08540, USA
Phone: +1 609 951 2992
Email: arup@ccrl.nj.nec.com

Rajiv Dighe
C&C Research Labs, NEC USA
4 Independence Way, Princeton
NJ 08540, USA
Phone: +1 609 951 2982
Email: rsd@ccrl.nj.nec.com

Furquan Ansari
C&C Research Labs, NEC USA
4 Independence Way, Princeton
NJ 08540, USA
Phone: +1 609 951 2965
Email: furquan@ccrl.nj.nec.com