Вы находитесь на странице: 1из 21

Multicast in L3VPNs


Bruce Davie1 bsd@cisco.com

1. Not a draft co-author, or a multicast expert

Aiming to encourage more involvement in multicast L3VPN work by providing user-friendly overview of problem space  Focus more on problems that the current proposed solutions  Hard questions will be deflected to draft authors

Likewise questions such as why did you choose to design it that way?

Attempts to make me look ignorant will be frowned upon


Recap of current (deployed) state (draft-rosen1)

See also draft-raggarwa-l3vpn-2547-mvpn-00.txt, draft-ycai-mboned-mvpn-deploy-00.txt

Enhancements/changes in L3VPN WG draft

Supporting multiple tree types Aggregation Carrying customer multicast routing in BGP Inter-AS improvements

Note Well: WG draft is a superset of draft-rosen

i.e. deployed solutions will not be obsoleted by new draft

L3VPN MulticastMotivation

Customers with IP multicast traffic would like to use MPLS VPN services RFC 2547/4364 only addresses unicast As usual, multicast makes the problem harder
Difficult to achieve same scalability as unicast

Multicast VPNCurrent Deployments


Based on draft-rosen-vpn-mcast-08.txt Many similarities to unicast, and some differences CE-routers maintain PIM adjacency with PE-router only
Similar concept to 2547/4364 VPNs

P-routers do not hold (S, G) state for individual customers

Unlike unicast, there is some per-customer state in P-routers

PE-routers exchange customer routing information using PIM

Contrast to BGP for unicast

Customer multicast group addresses need not be unique

Same as 2547/4364

Multicast VPNCurrent State (2)

Multicast domain is a set of multicast-enabled VRFs (mVRFs) that can send multicast traffic to each other
e.g., VRFs associated with a single customer

Maps all (S, G) that can exist in a particular VPN to a single (S, G) group in the P-network
This is the Multicast Distribution Tree (MDT) Amount of P-state is a function of # of VPNs rather than # of (S, G)s of all customers This is not as good as unicast, but better than the alternative

Mapping is achieved by encapsulating C-packet into P-packet using GRE

Default Multicast Distribution Tree

Customer B Default MDT PE PE PE

Customer B CE

Customer B CE

Customer B CE


PE routers build a default MDT in the global table for each mVRF using standard PIM procedures All PEs participating in the same mVPN join the same Default MDT Every mVRF must have a Default MDT MDT group addresses are defined by the provider
Unrelated to the groups used by the customer

Data MDTs may be created for high BW sources

Default Multicast Distribution Tree

Root Leaf PE PE Multicast Tunnel Interface Customer B Default MDT PE

Customer B CE

Customer B CE

Customer B CE


Default MDT is used as a permanent channel for PIM control messages and low bandwidth streams Access to the Default MDT is via a Multicast Tunnel Interface A PE is always a root (source) of the MDT A PE is also a leaf (receiver) to the MDT rooted on remote PEs

Limitations of draft-rosen

At least one multicast tree per customer in core

No option to aggregate multicast customers on one tree

Multicast traffic is GRE (not MPLS) encapsulated

Bandwidth and encaps/decaps cost Operational costdifferent mcast and unicast data planes


PIM the only fully described way to build core trees

Cant leverage p2mp RSVP-TE, mLDP

PE-PE exchange of C-routes using per-customer PIM instances  Inter-AS challenges

PMSI: P-Multicast Service Interface

New terms introduced to decouple tree from service  A PE delivers packet to PMSI, then all the required PEs will get than packet and known which MVPN it belongs to  Three types of PMSI

MI-PMSI: Multipoint Inclusive, all

all all

All PEs of an MVPN can transmit to all PEs

UI-PMSI: Unidirectional Inclusive, some Selective: S-PMSI, some some

Unidirectional, selected PEs can transmit to all PEs Unidirectional, selected PEs can transmit to selected PEs

Types of Multicast Trees


Inclusive: contains all the PEs for a given MVPN Selective: contains only a subset of PEs of a given MVPN Aggregate
Carries traffic for more than one MVPN May be either inclusive or selective

Supporting Multiple Tree Types

A PMSI is scoped to a single MVPN  PMSI is instantiated using a tunnel (or set of tunnels)  Tunnels may be:

PIM (any flavor) MPLS (mLDP or p2mp RSVP-TE) Unicast tunnels with ingress PE replication

Can map multiple PMSIs onto one tunnel (aggregation)  Encaps a function of tunnel, not service  Single provider can mix and match tunnel types

Mappings to Old Terminology

Default MDT
MI-PMSI, instantiated by PIM Shared Tree or set of PIM Source Trees


Data MDT
S-PMSI, instantiated by PIM Source Tree

New terminology helpful in:

Describing the complete set of options Allowing multiple instantiations of same service, without changing service spec

The process of discovering all the PEs with members in a given MVPN  Similar to RFC4364, but:

New address family MCAST-VPN Contains address of the originating PE Contains tunnel attribute to specify what sort of tunnel (e.g. PIM-SSM, mLDP, etc.) can be supported by this PE, and whether aggregation is supported
May contain a tunnel ID

Can also be used to discover set of PEs interested in a given group (to enable S-PMSI creation)


Conflicting goals
Scale: Minimize P-router state Use as few trees as possible Optimality: Send traffic at most once on each link, and only to PEs that need it Use a tree for each customer multicast group

Solution: lots of options

Draft-rosen has one MDT per VPN, and optional data MDT for high BW or sparse customer groups New draft also allows a tunnel to be shared among multiple mVPNs
Better aggregation, less optimality Requires a de-multiplexing field (e.g., MPLS label)

Utility of aggregation depends on amount of congruence and traffic volume

PE-PE Routing Exchange

In draft-rosen, C-PIM instances exchange PIM messages over the MDT as if it were a LAN
Per-customer PIM peering among the PEs By contrast, one BGP instance carries all customer unicast routes among PEs PIM Hellos can be eliminated, but Join/Prunes remain

In new draft, BGP is proposed, as in unicast

New AFI/SAFI Advertisement contains essentially the same info as a PIM join or prune (source, group, PE sending the message) RDs are used to disambiguate customer multicast group and source addresses BGP route reflectors may be used


Old (draft-rosen) approach: tunnel spans multiple ASes

Undesirable fate-sharing, must agree on tunnel type

New draft allows another approach: may splice tunnels from different ASes
Allows each AS to build its tunnels independently of other ASes Scaling now independent of number of PEs in other ASes

Inter-AS Overlay Tunnel


For a given MVPN, each AS independently builds an intra-AS tunnel In addition, an overlay tunnel that spans multiple ASes is built Each PE announces its MVPN membership info to the ASBRs with iBGP An ASBR announces the AS MVPN membership to other ASBRs (in other ASes) using eBGP This enables an AS-level spanning tree to be built among the set of ASes with members in this MVPN
Inter-AS tunnels spliced to intra-AS tunnels

Inter-AS Tunnels
Customer A Customer A



Customer A ASBR3

Intra-AS tunnels Inter-AS tunnels

Customer A

Other issues

RPF establishment
PEs need to know who their RPF PE is for a given route

Duplicate detection
Multihomed sites and switching from shared to source tree can lead to a PE getting duplicate packets draft proposes means to address this

C-RP Engineering
RP location in customer sites may cause hairpin PEs may act as outsourced C-RPs PEs may speak MSDP to C-RPs


WG draft builds on rosen draft without obsoleting it:

Support for more tree types, including PIM variants, mLDP, RSVP-TE Separation of service and mechanism Aggregation support More Inter-AS options with improved independence Possibility to use BGP for C-route exchange