PIM flooding mechanism and source discoveryCisco Systems, Inc.De kleetlaan 6aDiegem1831Belgiumice@cisco.comCisco Systems, Inc.Tasman DriveSan JoseCA 95134USAstig@cisco.comAegis BMD Program Office17211 Avenue D, Suite 160DahlgrenVA 22448-5148USAmichael.brig@mda.milSwedish Defence Material Administration (FMV)Lönnvägen 4Växjö35243Swedenanders@jomac.se
Routing
Multicast
PIM Sparse-Mode uses a Rendezvous Point and shared trees to forward
multicast packets from new sources. Once last hop routers receive
packets from a new source, they may join the Shortest Path Tree for
the source for optimal forwarding. This draft defines a new mechanism
that provides a way to support PIM Sparse Mode (SM) without the need for
PIM registers, RPs or shared trees. Multicast source information is flooded
throughout the multicast domain using a new generic PIM flooding mechanism.
This allows last hop routers to learn about new sources without receiving
initial data packets.
PIM Sparse-Mode uses a Rendezvous Point (RP) and shared trees to forward
multicast packets to Last Hop Routers (LHR). After the first packet is
received by a LHR, the source of the multicast stream is learned and the
Shortest Path Tree (SPT) can be joined. This draft defines a new mechanism
that
provides a way to support PIM Sparse Mode (SM) without the need for PIM
registers, RPs or shared trees. Multicast source information is flooded
throughout the multicast domain using a new generic PIM flooding mechanism.
By removing the need for
RPs and shared trees, the PIM-SM procedures are simplified, improving router
operations, management and making the protocol more robust. Also the data
packets are only sent on the SPTs, providing optimal forwarding.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL
NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described
in RFC 2119 .
Rendezvous Point
Bootstrap Router
Reverse Path Forwarding
Shortest Path Tree
First Hop Router, directly connected to the source
Last Hop Router, directly connected to the receiver
PIM Flooding Mechanism
PFM Source Announcement
Multicast source to group mapping
A prototype of this specification has been implemented and there has
been some limited testing in the field.
The prototype was tested in a network with low bandwidth radio links.
The network has frequent topology changes, including frequest link
or router failures. Previously existing mechanisms like PIM-SM and
PIM-DM were tested.
With PIM-SM the existing RP election mechanisms were found to be too
slow. With PIM-DM, issues were observed with new multicast sources
starving low bandwidth links even when there are no receivers, in
some cases such
that there was no bandwidth left for prune message.
For the PFM-SA prototype tests, all routers were configured to send
PFM-SA for directly
connected source and to cache received announcements. Applications such
as SIP with multicast subscriber discovery, multicast voice
conferencing, position tracking and NTP were successfully tested. The
tests went quite well. Packets were rerouted as needed and there were no
unnecessary forwarding of packets. Ease of configuration was seen as
a plus.
The Bootstrap Router mechanism (BSR) is a commonly used
mechanism for distributing dynamic Group to RP mappings in PIM. It is
responsible for flooding information about such mappings throughout a PIM
domain, so that all routers in the domain can have the same information.
BSR as defined, is only able to distribute Group to RP mappings. We are
defining a more generic mechanism that can flood any kind of information
throughout a PIM domain. It is not necessarily a domain though, it depends
on the administrative boundaries being configured. The forwarding rules are
identical to BSR, except that one can
control whether routers should forward unsupported data types. For
some types of information it is quite useful that it can be distributed
without all routers having to support the particular type, while there may
also be types where it is necessary for every single router to support it.
The mechanism includes an originator address which is used for RPF checking
to restrict the flooding, and prevent loops, just like BSR. Like BSR,
messages are forwarded hop by hop. Note that there is no equivalent to the
BSR election mechanism;, there can be multiple originators. We call this
mechanism the PIM Flooding Mechanism (PFM).
Reserved, Checksum Described in .
PIM Message Type. Value (pending IANA) for a PFM message.
When set, this bit means that the PFM message is not to be forwarded.
The address of the router that originated the message. This can be any
address assigned to the originating router, but MUST be routable in the domain
to allow successful forwarding. The format for this
address is given in the Encoded-Unicast address in .
A message contains one or more TLVs, in this case n TLVs. The Type
specifies what kind of information is in the Value.
The length of the the value field.
The value associated with the type and of the specified length.
A router that receives a PFM message MUST perform the initial checks
specified here. If the checks fail, the message MUST be dropped. An error
MAY be logged, but otherwise the message MUST be dropped silently. If the
checks pass, the contents is processed according to the processing rules
of the included TLVs.
In order to do further processing, a message MUST meet the following
requirements. The message MUST be from a directly connected
neighbor for which we have active Hello state, and it MUST have been
sent to the ALL-PIM-ROUTERS group. Also, the interface
MUST NOT be an administrative boundary for PFM.
If No-Forward is not set, it MUST have been sent by the RPF neighbor
for the originator address. If No-Forward is set,
we MUST have restarted within 60 seconds.
In pseudo-code the algorithm is as follows:
Note that src_ip_address is the source address in the IP header of
the PFM message. Originator is the originator field inside
the PFM message, and is the router that originated the message.
When the message is forwarded hop by hop, the originator address
never changes, while the source address will be an address belonging
to the router that last forwarded the message.
When the message is received, the initial checks above must be
performed. If it passes the checks, we then for each included
TLV perform processing according to the specification for that
TLV.
After processing, we forward the message. Unless otherwise
specified by the type specification, the TLVs in the forwarded
message are identical to the TLVs in the received message.
However, if the most significant bit in the type field is set (the
type value is larger than 32767) and we do not support the type,
then that particular type should be omitted from the forwarded
messages. The message is forwarded out of all interfaces with PIM
neighbors (including the interface it was received on).
The generic flooding mechanism (PFM) defined in the previous section
can be used for distributing source to group mappings about active
multicast sources throughout a PIM domain. A Group Source Holtime (GSH)
TLV is defined for this purpose.
This TLV has type 0.
The length of the value.
The group we are announcing sources for. The format
for this address is given in the Encoded-Group format in
.
How many unicast encoded sources address encodings follow.
The Holdtime (in seconds) for the corresponding source(s).
The source address for the corresponding group.
The format for these addresses is given in the Encoded-Unicast
address in .
A PFM message MAY contain one or more
Group Source Holdtime (GSH) TLVs. This is used to flood information about
active multicast sources. Each FHR that is directly connected to an active
multicast source originates PFM messages containing GSH TLVs.
How a multicast router discovers
the source of the multicast packet and when it considers itself the FHR
follows the same procedures as the registering process described in
. When a FHR has decided that a register needs to be
sent per , the SG is not registered via the PIM SM
register procedures, but the SG mapping is included in an GSH TLV in a PFM
message. Note, only the SG mapping is distributed in the message, not the
entire packet as
would have been done with a PIM register. The router originating the PFM
messages includes one of its own addresses in the originator field. Note that
this address SHOULD be routeable due to RPF checking. The PFM messages
containing the GSH TLV are periodically sent for as long as the multicast
source is active, similar to
how PIM registers are periodically sent. The default announcement period is
60 seconds, which means that as long as the source is active, it is included
in a PFM message originated every 60 seconds. The holdtime for the source is
by default 210 seconds. Other values MAY be configured, but the holdtime MUST
be either zero, or larger than the announcement period. It is RECOMMENDED to
be 3.5 times the announcement period. A source MAY be announced with a
holdtime of zero to indicate that the source is no longer active.
If an implementation supports originating GSH TLVs with different
holdtimes for different sources, it can if needed send multiple TLVs
with the same group address. Due to the format, all the sources in
the same TLV have the same holdtime.
A router that receives a PFM message containing GSH TLVs SHOULD parse the
message and store each of the GSH TLVs as SG mappings with a holdtimer started
with the advertised holdtime.
For each group that has directly connected receivers, this router
SHOULD send PIM (S,G) joins for all the SG mappings advertised in the message
for the group. The SG mappings are kept alive for as long as the holdtimer for
the source is running. Once the holdtimer expires a PIM router MAY send a
PIM (S,G) prune to remove itself from the tree. However, when this happens,
there should be no more packets sent by the source, so it may be desirable
to allow the state to time out rather than sending a prune.
Note that a holdtime of zero
has a special meaning. It is to be treated as if the source just expired,
and state to be removed. Source information MUST NOT be removed due to the
source being omitted in a message. For instance, if there
is a large number of sources for a group, there may be multiple PFM messages,
each message containing a different list of sources for the group.
The PIM register procedure is designed to deliver Multicast packets to the RP
in the absence of a Shortest Path Tree (SPT) from the RP to the source. The
register packets received on the RP are decapsulated and forwarded down the
shared tree to the LHRs. As soon as an SPT is built, multicast packets would
flow natively over the SPT to the RP or LHR and the register process would stop.
The PIM register process ensures packet delivery until an SPT is in place
reaching the FHR. If the packets were not unicast encapsulated to the RP they
would be dropped by the FHR until the SPT is setup. This functionality is
important for applications where the initial packet(s) must be received for
the application to work correctly. Another reason would be for bursty sources.
If the application sends out a multicast packet every 4 minutes (or longer),
the SPT is torn down (typically after 3:30 minutes of inactivity) before the
next packet is forwarded down the tree. This will cause no multicast packet
to ever be forwarded. A well behaved application should be able to deal
with packet loss since IP is a best effort based packet delivery system.
But in reality this is not always the case.
With the procedures defined in this document the packet(s) received by the FHR
will be dropped until the LHR has learned about the source and the SPT is
built. That means for bursty sources or applications sensitive for the delivery
of the first packet this solution would not be very applicable. This solution
is mostly useful for applications that don't have strong dependency on the
initial packet(s) and have a fairly constant data rate, like video distribution
for example. For applications with strong dependency on the initial packet(s)
we recommend using PIM Bidir or SSM
. The protocol operations are much simpler compared to
PIM SM, it will cause less churn in the network and both guarantee best effort
delivery for the initial packet(s).
In a PIM SM deployment where the network becomes partitioned, due to link
or node failure, it is possible that the RP becomes unreachable to a certain
part of the network. New sources that become active in that partition will not
be able to register to the RP and receivers within that partition are not able
to receive the traffic. Ideally you would want to have a candidate RP in each
partition, but you never know in advance which routers will form a partitioned
network. In order to be fully resilient, each router in the network may end up
being a candidate RP. This would increase the operational complexity of the
network.
The solution described in this document does not suffer from that problem. If
a network becomes partitioned and new sources become active, the receivers in
that partitioned will receive the SG Mappings and join the source tree. Each
partition works independently of the other partition(s) and will continue to
have access to sources within that partition. As soon as the network heals,
the SG Mappings are re-flooded into the other partition(s) and other receivers
can join to the newly learned sources.
The security considerations are mainly similar to what is documented in
. It is a concern that rogue devices can inject
packets that are flooded throughout a domain. PFM packets must only be
accepted from a PIM neighbor. Deployments may use mechanisms for
authenticating PIM neighbors. For PFM-SA it is an issue that injected
packets from a rogue device could send SG mappings for a large number of
source addresses, causing routers to use memory storing these mappings, and
also if they have interest in the groups, build Shortest Path Trees for
sources that are not actually active.
This document requires the assignment of a new PIM message type for the
PIM Flooding Mechanism (PFM). IANA is also
requested to create a registry for PFM TLVs, with type 0 assigned to the
"Source Group Holdtime" TLV. Values in the range 1-65535 are "Unassigned".
Assignments for the registry are to be made
according to the policy "IETF Review" as defined in .
The authors would like to thank Arjen Boers for contributing to the
initial idea, and Yiqun Cai and Dino Farinacci for their comments on the draft.