Network Working Group G. Armitage
Request for Comments: 2191 LUCent Technologies
Category: Informational September 1997
VENUS - Very Extensive Non-Unicast Service
Status of this Memo
This memo provides information for the Internet community. This memo
does not specify an Internet standard of any kind. Distribution of
this memo is unlimited.
Abstract
The MARS model (RFC2022) provides a solution to intra-LIS IP
multicasting over ATM, establishing and managing the use of ATM pt-
mpt SVCs for IP multicast packet forwarding. Inter-LIS multicast
forwarding is achieved using Mrouters, in a similar manner to which
the "Classical IP over ATM" model uses Routers to inter-connect LISes
for unicast traffic. The development of unicast IP shortcut
mechanisms (e.g. NHRP) has led some people to request the
development of a Multicast equivalent. There are a number of
different approaches. This document focuses exclusively on the
problems associated with extending the MARS model to cover multiple
clusters or clusters spanning more than one subnet. It describes a
hypothetical solution, dubbed "Very Extensive NonUnicast Service"
(VENUS), and shows how complex such a service would be. It is also
noted that VENUS ultimately has the look and feel of a single, large
cluster using a distributed MARS. This document is being issued to
help focus ION efforts towards alternative solutions for establishing
ATM level multicast connections between LISes.
1. Introduction
The classical model of the Internet running over an ATM cloud
consists of multiple Logical IP Subnets (LISs) interconnected by IP
Routers [1]. The evolving IP Multicast over ATM solution (the "MARS
model" [2]) retains the classical model. The LIS becomes a "MARS
Cluster", and Clusters are interconnected by conventional IP
Multicast routers (Mrouters).
The development of NHRP [3], a protocol for discovering and managing
unicast forwarding paths that bypass IP routers, has led to some
calls for an IP multicast equivalent. Unfortunately, the IP
multicast service is a rather different beast to the IP unicast
service. This document aims to eXPlain how much of what has been
learned during the development of NHRP must be carefully scrutinized
before being re-applied to the multicast scenario. Indeed, the
service provided by the MARS and MARS Clients in [2] are almost
orthogonal to the IP unicast service over ATM.
For the sake of discussion, let"s call this hypothetical multicast
shortcut discovery protocol the "Very Extensive Non-Unicast Service"
(VENUS). A "VENUS Domain" is defined as the set of hosts from two or
more participating Logical IP Subnets (LISs). A multicast shortcut
connection is a point to multipoint SVC whose leaf nodes are
scattered around the VENUS Domain. (It will be noted in section 2
that a VENUS Domain might consist of a single MARS Cluster spanning
multiple LISs, or multiple MARS Clusters.)
VENUS faces a number of fundamental problems. The first is exploding
the scope over which individual IP/ATM interfaces must track and
react to IP multicast group membership changes. Under the classical
IP routing model Mrouters act as aggregation points for multicast
traffic flows in and out of Clusters [4]. They also act as
aggregators of group membership change information - only the IP/ATM
interfaces within each Cluster need to know the specific identities
of their local (intra-cluster) group members at any given time.
However, once you have sources within a VENUS Domain establishing
shortcut connections the data and signaling plane aggregation of
Mrouters is lost. In order for all possible sources throughout a
VENUS Domain to manage their outgoing pt-mpt SVCs they must be kept
aware of MARS_JOINs and MARS_LEAVEs occuring in every MARS Cluster
that makes up a VENUS Domain. The nett effect is that a VENUS domain
looks very similar to a single, large distributed MARS Cluster.
A second problem is the impact that shortcut connections will have on
IP level Inter Domain Multicast Routing (IDMR) protocols. Multicast
groups have many sources and many destinations scattered amongst the
participating Clusters. IDMR protocols assume that they can calculate
efficient inter-Cluster multicast trees by aggregating individual
sources or group members in any given Cluster (subnet) behind the
Mrouter serving that Cluster. If sources are able to simply bypass an
Mrouter we introduce a requirement that the existence of each and
every shortcut connection be propagated into the IDMR decision making
processes. The IDMR protocols may need to adapt when a source"s
traffic bypasses its local Mrouter(s) and is injected into Mrouters
at more distant points on the IP-level multicast distribution tree.
(This issue has been looked at in [7], focussing on building
forwarding trees within networks where the termination points are
small in number and sparsely distributed. VENUS introduces tougher
requirements by assuming that multicast group membership may be dense
across the region of interest.)
This document will focus primarily on the internal problems of a
VENUS Domain, and leave the IDMR interactions for future analysis.
2. What does it mean to "shortcut" ?
Before going further it is worth considering both the definition of
the Cluster, and two possible definitions of "shortcut".
2.1 What is a Cluster?
In [2] a MARS Cluster is defined as the set of IP/ATM interfaces that
are willing to engage in direct, ATM level pt-mpt SVCs to perform IP
multicast packet forwarding. Each IP/ATM interface (a MARS Client)
must keep state information regarding the ATM addresses of each leaf
node (recipient) of each pt-mpt SVC it has open. In addition, each
MARS Client receives MARS_JOIN and MARS_LEAVE messages from the MARS
whenever there is a requirement that Clients around the Cluster need
to update their pt-mpt SVCs for a given IP multicast group.
It is worth noting that no MARS Client has any concept of how big its
local cluster is - this knowledge is kept only by the MARS that a
given Client is registered with.
Fundamentally the Cluster (and the MARS model as a whole) is a
response to the requirement that any multicast IP/ATM interface using
pt-mpt SVCs must, as group membership changes, add and drop leaf
nodes itself. This means that some mechanism, spanning all possible
group members within the scopes of these pt-mpt SVCs, is required to
collect group membership information and distribute it in a timely
fashion to those interfaces. This is the MARS Cluster, with certain
scaling limits described in [4].
2.2 LIS/Cluster boundary "shortcut"
The currently popular definition of "shortcut" is based on the
existence of unicast LIS boundaries. It is tied to the notion that
LIS boundaries have physical routers, and cutting through a LIS
boundary means bypassing a router. Intelligently bypassing routers
that sit at the edges of LISs has been the goal of NHRP. Discovering
the ATM level identity of an IP endpoint in a different LIS allows a
direct SVC to be established, thus shortcutting the logical IP
topology (and very real routers) along the unicast path from source
to destination.
For simplicity of early adoption RFC2022 recommends that a Cluster"s
scope be made equivalent to that of a LIS. Under these circumstances
the "Classical IP" routing model places Mrouters at LIS/Cluster
boundaries, and multicast shortcutting must involve bypassing the
same physical routing entities as unicast shortcutting. Each MARS
Cluster would be independent and contain only those IP/ATM interfaces
that had been assigned to the same LIS.
As a consequence, a VENUS Domain covering the hosts in a number of
LIS/Clusters would have to co-ordinate each individual MARS from each
LIS/Cluster (to ensure group membership updates from around the VENUS
Domain were propagated correctly).
2.3 Big Cluster, LIS boundary "shortcut"
The MARS model"s fundamental definition of a Cluster was deliberately
created to be independent of unicast terminology. Although not
currently well understood, it is possible to build a single MARS
Cluster that encompasses the members of multiple LISs. As expected,
inter-LIS unicast traffic would pass through (or bypass, if using
NHRP) routers on the LIS boundaries. Also as expected, each IP/ATM
interface, acting as a MARS Client, would forward their IP multicast
packets directly to intra-cluster group members. However, because the
direct intra-cluster SVCs would exist between hosts from the
different LISs making up the cluster, this could be considered a
"shortcut" of the unicast LIS boundaries.
This approach immediately brings up the problem of how the IDMR
protocols will react. Mrouters only need to exist at the edges of
Clusters. In the case of a single Cluster spanning multiple LISs,
each LIS becomes hidden behind the Mrouter at the Cluster"s edge.
This is arguably not a big problem if the Cluster is a stub on an
IDMR protocol"s multicast distribution tree, and if there is only a
single Mrouter in or out of the Cluster. Problems arise when two or
more Mrouters are attached to the edges of the Cluster, and the
Cluster is used for transit multicast traffic. Each Mrouter"s
interface is assigned a unicast identity (e.g. that of the unicast
router containing the Mrouter). IDMR protocols that filter packets
based on the correctness of the upstream source may be confused at
receiving IP multicast packets directly from another Mrouter in the
same cluster but notionally "belonging" to an LIS multiple unicast IP
hops away.
Adjusting the packet filtering algorithms of Mrouters is something
that needs to be addressed by any multicast shortcut scheme. It has
been noted before and a solution proposed in [7]. For the sake of
argument this document will assume the problem solvable. (However, it
is important that any solution scales well under general topologies
and group membership densities.)
A multi-LIS MARS Cluster can be considered a simple VENUS Domain.
Since it is a single Cluster it can be scaled using the distributed
MARS solutions currently being developed within the IETF [5,6].
3. So what must VENUS look like?
A number of functions that occur in the MARS model are fundamental to
the problem of managing root controlled, pt-mpt SVCs. The initial
setup of the forwarding SVC by any one MARS Client requires a
query/response exchange with the Client"s local MARS, establishing
who the current group members are (i.e. what leaf nodes should be on
the SVC). Following SVC establishment comes the management phase -
MARS Clients need to be kept informed of group membership changes
within the scopes of their SVCs, so that leaf nodes may be added or
dropped as appropriate.
For intra-cluster multicasting the current MARS approach is our
solution for these two phases.
For the rest of this document we will focus on what VENUS would look
like when a VENUS Domain spans multiple MARS Clusters. Under such
circumstances VENUS is a mechanism co-ordinating the MARS entities of
each participating cluster. Each MARS is kept up to date with
sufficient domain-wide information to support both phases of client
operation (SVC establishment and SVC management) when the SVC"s
endpoints are outside the immediate scope of a client"s local MARS.
Inside a VENUS Domain a MARS Client is supplied information on group
members from all participating clusters.
The following subsections look at the problems associated with both
of these phases independently. To a first approximation the problems
identified are independent of the possible inter-MARS mechanisms. The
reader may assume the MARS in any cluster has some undefined
mechanism for communicating with the MARSs of clusters immediately
adjacent to its own cluster (i.e. connected by a single Mrouter hop).
3.1 SVC establishment - answering a MARS_REQUEST.
The SVC establishment phase contains a number of inter-related
problems.
First, the target of a MARS_REQUEST (an IP multicast group) is an
abstract entity. Let us assume that VENUS does not require every MARS
to know the entire list of group members across the participating
clusters. In this case each time a MARS_REQUEST is received by a
MARS from a local client, the MARS must construct a sequence of
MARS_MULTIs based on locally held information (on intra-cluster
members) and remotely solicited information.
So how does it solicit this information? Unlike the unicast
situation, there is no definite, single direction to route a
MARS_REQUEST across the participating clusters. The only "right"
approach is to send the MARS_REQUEST to all clusters, since group
members may exist anywhere and everywhere. Let us allow one obvious
optimization - the MARS_REQUEST is propagated along the IP multicast
forwarding tree that has been established for the target group by
whatever IDMR protocol is running at the time.
As noted in [4] there are various reasons why a Cluster"s scope be
kept limited. Some of these (MARS Client or ATM NIC limitations)
imply that the VENUS discovery process not return more group members
in the MARS_MULTIs that the requesting MARS Client can handle. This
provides VENUS with an interesting problem of propagating out the
original MARS_REQUEST, but curtailing the MARS_REQUESTs propagation
when a sufficient number of group members have been identified.
Viewed from a different perspective, this means that the scope of
shortcut achievable by any given MARS Client may depend greatly on
the shape of the IP forwarding tree away from its location (and the
density of group members within clusters along the tree) at the time
the request was issued.
How might we limit the number of group members returned to a given
MARS Client? Adding a limit TLV to the MARS_REQUEST itself is
trivial. At first glance it might appear that when the limit is being
reached we could summarize the next cluster along the tree by the ATM
address of the Mrouter into that cluster. The nett effect would be
that the MARS Client establishes a shortcut to many hosts that are
inside closer clusters, and passes its traffic to more distant
clusters through the distant Mrouter. However, this approach only
works passably well for a very simplistic multicast topology (e.g. a
linear concatenation of clusters).
In a more general topology the IP multicast forwarding tree away from
the requesting MARS Client will branch a number of times, requiring
the MARS_REQUEST to be replicated along each branch. Ensuring that
the total number of returned group members does not exceed the
client"s limit becomes rather more difficult to do efficiently.
(VENUS could simply halve the limit value each time it split a
MARS_REQUEST, but this might cause group member discovery on one
branch to end prematurely while all the group members along another
branch are discovered without reaching the subdivided limit.)
Now consider this decision making process scattered across all the
clients in all participating clusters. Clients may have different
limits on how many group members they can handle - leading to
situations where different sources can shortcut to different
(sub)sets of the group members scattered across the participating
clusters (because the IP multicast forwarding trees from senders in
different clusters may result in different discovery paths being
taken by their MARS_REQUESTs.)
Finally, when the MARS_REQUEST passes a cluster where the target
group is MCS supported, VENUS must ensure the ATM address of the MCS
is collected rather than the addresses of the actual group members.
(To do otherwise would violate the remote cluster"s intra-cluster
decision to use an MCS. The shortcut in this case must be content to
directly reach the remote cluster"s MCS.)
(A solution to part of this problem would be to ensure that a VENUS
Domain never has more MARS Clients throughout than the clients are
capable of adding as leaf nodes. This may or may not appeal to
people"s desire for generality of a VENUS solution. It also would
appear to beg the question of why the problem of multiple-LIS
multicasting isn"t solved simply by creating a single big MARS
Cluster.)
3.2 SVC management - tracking group membership changes.
Once a client"s pt-mpt SVC is established, it must be kept up to
date. The consequence of this is simple, and potentially
devastating: The MARS_JOINs and MARS_LEAVEs from every MARS Client in
every participating cluster must be propagated to every possible
sender in every participating cluster (this applies to groups that
are VC Mesh supported - groups that are MCS supported in some or all
participating clusters introduce complications described below).
Unfortunately, the consequential signaling load (as all the
participating MARSs start broadcasting their MARS_JOIN/LEAVE
activity) is not localized to clusters containing MARS Clients who
have established shortcut SVCs. Since the IP multicast model is Any
to Multipoint, and you can never know where there may be source MARS
Clients, the JOINs and LEAVEs must be propagated everywhere, always,
just in case. (This is simply a larger scale version of sending JOINs
and LEAVEs to every cluster member over ClusterControlVC, and for
exactly the same reason.)
The use of MCss in some clusters instead of VC Meshes significantly
complicates the situation, as does the initial scoping of a client"s
shortcut during the SVC establishment phase (described in the
preceding section).
In Clusters where MCSs are supporting certain groups, MARS_JOINs or
MARS_LEAVEs are only propagated to MARS Clients when an MCS comes or
goes. However, it is not clear how to effectively accommodate the
current MARS_MIGRATE functionality (that allows a previously VC Mesh
based group to be shifted to an MCS within the scope of a single
cluster). If an MCS starts up within a single Cluster, it is possible
to shift all the intra-cluster senders to the MCS using MARS_MIGRATE
as currently described in the MARS model. However, MARS Clients in
remote clusters that have shortcut SVCs into the local cluster also
need some signal to shift (otherwise they will continue to send their
packets directly to the group members in the local cluster).
This is a non-trivial requirement, since we only want to force the
remote MARS Clients to drop some of their leaf nodes (the ones to
clients within the Cluster that now has an MCS), add the new MCS as a
leaf node, and leave all their other leaf nodes untouched (the cut-
through connections to other clusters). Simply broadcasting the
MARS_MIGRATE around all participating clusters would certainly not
work. VENUS needs a new control message with semantics of "replaced
leaf nodes {x, y, z} with leaf node {a}, and leave the rest alone".
Such a message is easy to define, but harder to use.
Another issue for SVC management is that the scope over which a MARS
Client needs to receive JOINs and LEAVEs needs to respect the
Client"s limited capacity for handling leaf nodes on its SVC. If the
MARS Client initially issued a MARS_REQUEST and indicated it could
handle 1000 leaf nodes, it is not clear how to ensure that subsequent
joins of new members wont exceed that limit. Furthermore, if the SVC
establishment phase decided that the SVC would stop at a particular
Mrouter (due to leaf node limits being reached), the Client probably
should not be receiving direct MARS_JOIN or MARS_LEAVE messages
pertaining to activity in the cluster "behind" this Mrouter. (To do
otherwise could lead to multiple copies of the source client"s
packets reaching group members inside the remote cluster - one
version through the Mrouter, and another on the direct SVC connection
that the source client would establish after receiving a subsequent,
global MARS_JOIN regarding a host inside the remote cluster.)
Another scenario involves the density of group members along the IDMR
multicast tree increasing with time after the initial MARS_REQUEST is
answered. Subsequent JOINs from Cluster members may dictate that a
"closer" Mrouter be used to aggregate the source"s outbound traffic
(so as not to exceed the source"s leaf node limitations). How to
dynamically shift between terminating on hosts within a Cluster, and
terminating on a cluster"s edge Mrouter, is an open question.
To complicate matters further, this scoping of the VENUS domain-wide
propagation of MARS_JOINs and MARS_LEAVEs needs to be on a per-
source- cluster basis, at least. If MARS Clients within the same
cluster have different leaf node limits, the problem worsens. Under
such circumstances, one client may have been able to establish a
shortcut SVC directly into a remote cluster while a second client -
in the same source cluster - may have been forced to terminate its
shortcut on the remote cluster"s Mrouter. The first client obviously
needs to know about group membership changes in the remote cluster,
whilst the second client does not. Propagating these JOIN/LEAVE
messages on ClusterControlVC in the source cluster will not work -
the MARS for the source cluster will need to explicitly send copies
of the JOIN/LEAVE messages only to those MARS Clients whose prior SVC
establishment phase indicates they need them. Propagation of messages
to indicate a VC Mesh to MCS transition within clusters may also need
to take account of the leaf node limitations of MARS Clients. The
scaling characteristics of this problem are left to the readers
imagination.
It was noted in the previous section that a VENUS domain could be
limited to ensure there are never more MARS Clients than any one
client"s leaf node limit. This would certainly avoid the need to for
complicated MARS_JOIN/LEAVE propagation mechanisms. However, it begs
the question of how different the VENUS domain then becomes from a
single, large MARS Cluster.
4. What is the value in bypassing Mrouters?
This is a good question, since the whole aim of developing a shortcut
connection mechanism is predicated on the assumption that bypassing
IP level entities is always a "win". However, this is arguably not
true for multicast.
The most important observation that should be made about shortcut
connection scenarios is that they increase the exposure of any given
IP/ATM interface to externally generated SVCs. If there are a
potential 1000 senders in a VENUS Domain, then you (as a group
member) open yourself up to a potential demand for 1000 instances of
your re-assembly engine (and 1000 distinct incoming SVCs, when you
get added as a leaf node to each sender"s pt-mpt SVC, which your
local switch port must be able to support).
It should be no surprise that the ATM level scaling limits applicable
to a single MARS Cluster [4] will also apply to a VENUS Domain. Again
we"re up against the question of why you"d bypass an Mrouter. As
noted in [4] Mrouters perform a useful function of data path
aggregation - 100 senders in one cluster become 1 pt-mpt SVC out of
the Mrouter into the next cluster along the tree. They also hide MARS
signaling activity - individual group membership changes in one
cluster are hidden from IP/ATM interfaces in surrounding clusters.
The loss of these benefits must be factored into any network designed
to utilize multicast shortcut connections.
(For the sake of completeness, it must be noted that extremely poor
mismatches of IP and ATM topologies may make Mrouter bypass
attractive if it improves the use of the underlying ATM cloud. There
may also be benefits in removing the additional re-
assembly/segmentation latencies of having packets pass through an
Mrouter. However, a VENUS Domain ascertained to be small enough to
avoid the scaling limits in [4] might just as well be constructed as
a single large MARS Cluster. A large cluster also avoids a
topological mismatch between IP Mrouters and ATM switches.)
5. Relationship to Distributed MARS protocols.
The ION working group is looking closely at the development of
distributed MARS architectures. An outline of some issues is provided
in [5,6]. As noted earlier in this document the problem space looks
very similar that faced by our hypothetical VENUS Domain. For
example, in the load-sharing distributed MARS model:
- The Cluster is partitioned into sub-clusters.
- Each Active MARS is assigned a particular sub-cluster, and uses
its own sub-ClusterControlVC to propagate JOIN/LEAVE messages to
members of its sub-cluster.
- The MARS_REQUEST from any sub-cluster member must return
information from all the sub-clusters, so as to ensure that all a
group"s members across the cluster are identified.
- Group membership changes in any one sub-cluster must be
immediately propagated to all the other sub-clusters.
There is a clear analogy to be made between a distributed MARS
Cluster, and a VENUS Domain made up of multiple single-MARS Clusters.
The information that must be shared between sub-clusters in a
distributed MARS scenario is similar to the information that must be
shared between Clusters in a VENUS Domain.
The distributed MARS problem is slightly simpler than that faced by
VENUS:
- There are no Mrouters (IDMR nodes) within the scope of the
distributed Cluster.
- In a distributed MARS Cluster an MCS supported group uses the
same MCS across all the sub-clusters (unlike the VENUS Domain,
where complete generality makes it necessary to cope with mixtures
of MCS and VC Mesh based Clusters).
6. Conclusion.
This document has described a hypothetical multicast shortcut
connection scheme, dubbed "Very Extensive NonUnicast Service"
(VENUS). The two phases of multicast support - SVC establishment,
and SVC management - are shown to be essential whether the scope is a
Cluster or a wider VENUS Domain. It has been shown that once the
potential scope of a pt-mpt SVC at establishment phase has been
expanded, the scope of the SVC management mechanism must similarly be
expanded. This means timely tracking and propagation of group
membership changes across the entire scope of a VENUS Domain.
It has also been noted that there is little difference in result
between a VENUS Domain and a large MARS Cluster. Both suffer from the
same fundamental scaling limitations, and both can be arranged to
provide shortcut of unicast routing boundaries. However, a completely
general multi-cluster VENUS solution ends up being more complex. It
needs to deal with bypassed Mrouter boundaries, and dynamically
changing group membership densities along multicast distribution
trees established by the IDMR protocols in use.
No solutions have been presented. This document"s role is to provide
context for future developments.
Acknowledgment
This document was prepared while the author was with the
Internetworking Research group at Bellcore.
Security Considerations
This memo addresses specific scaling issues associated with the
extension of the MARS architecture beyond that described in RFC2022.
It is an Informational memo, and does not mandate any additional
protocol behaviors beyond those described in RFC2022. As such, the
security implications are no greater or less than the implications
inherent in RFC2022. Should enhancements to security be required,
they would need to be added as an extension to the base architecture
in RFC2022.
Author"s Address
Grenville Armitage
Bell Labs, Lucent Technologies.
101 Crawfords Corner Rd,
Holmdel, NJ, 07733
USA
EMail: gja@dnrc.bell-labs.com
References
[1] Laubach, M., "Classical IP and ARP over ATM", RFC1577, Hewlett-
Packard Laboratories, December 1993.
[2] Armitage, G., "Support for Multicast over UNI 3.0/3.1 based ATM
Networks.", Bellcore, RFC2022, November 1996.
[3] Luciani, J., et al, "NBMA Next Hop Resolution Protocol (NHRP)",
Work in Progress, February 1997.
[4] Armitage, G., "Issues affecting MARS Cluster Size", Bellcore, RFC
2121, March 1997.
[5] Armitage, G., "Redundant MARS architectures and SCSP", Bellcore,
Work in Progress, November 1996.
[6] Luciani, J., G. Armitage, J. Jalpern, "Server Cache
Synchronization Protocol (SCSP) - NBMA", Work in Progress, March 1997.
[7] Rekhter, Y., D. Farinacci, " Support for Sparse Mode PIM over
ATM", Cisco Systems, Work in Progress, April 1996.