Hub and Spoke Multi-Point LSPs for Scalable VPLS Architecture

Multiprotocol Label Switching (MPLS) and Generalized MPLS (GMPLS) provide a mechanism to set up point-to-multipoint (P2MP) LSPs which carry traffic from one ingress point (the root node) to several egress points (the leaf nodes), thus enabling multicast forwarding in an MPLS domain. However, there is no provision to provide a co-routed path back from the egress points (the leaves) to the ingress node (the root). The only way to do this is by configuring Unidirectional point-to-point LSPs from the leaves back to the root node. This entails configuring each leaf node with an LSP back to the root which could be a configuration and management nightmare if the number of leaves are large. Second, it can also not guarantee a co-routed path back from the leaf to the node, as the process of setting up the Unidirectional P2P path is independent from setting up the P2MP path.

This post introduces the concept of a hub-and-spoke multipoint (HSMP) LSP that allows traffic from the root to the leaves via a P2MP LSP and back from the leaves to the root via a unidirectional co-routed LSP. The proposed technique targets one-to-many applications that require reverse one-to-one traffic flow (thus many one-to-one in the reverse direction).

Consider the figure shown below.

PE1 is the ingress router for the HSMP LSP. The egress routers (leaves) are PE2, PE3 and PE4. As can be seen from the figure, PE1 creates a single copy of each packet arriving from the data source. This packet carries the MPLS label value L1. P1 is an ordinary Label Switch Router (LSR) that swaps the incoming label L1 with L2. Lets now focus on the packet forwarding process on the node P2.

For each packet belonging to the HSMP LSP, P2 makes three copies (just like how its done for a P2MP LSP), each of which is sent to PE2, PE3 and PE4 respectively. Packets arrive at P2 with MPLS label L2. As shown in the above figure, P2’s ILM contains an entry for label L2 saying that one copy of the packet should be sent out on interface if1 with label L3, another copy on interface if2 with label L4 and a last copy on interface if3 with label L5. Since the LSR P2 is replicating the MPLS packets its called the branch node.

An obvious advantage of this scheme is the bandwidth optimization. If we had used unidirectional P2P LSPs instead of an HSMP LSP, PE1 would have sent three copies of each packet that it had received from the data source and thus congesting the links PE1-P1 and P1-P2.

What sets an HSMP LSP apart from a regular P2MP LSP is the ability in the former to set up a path from the leaves back to the root. P2MP LSPs are unidirectional, so no traffic can flow from the leaf node routers to the ingress head end router along the P2MP LSP. In HSMP LSP, the leaf nodes can also send unidirectional traffic back to the root. This is shown in the figure below.

Because of the mechanisms defined in HSMP LSP, the branch node P2 advertises the same upstream label L1 for a given HSMP LSP to the nodes PE2, PE3 and PE4. It programs its ILM table as shown above, where it simply swaps L1 with L2 for all incoming MPLS packets and sends those towards P1. This way HSMP LSP is also able to provide a path back from the leaf nodes to the root node.

In the last post i had discussed issues that exist in VPLS. Lets see how HSMP LSP can solve them. I am using the same topology as was used there to demonstrate how HSMP LSP helps.

The figure 3 above shows the same VPLS service that we had discussed in the earlier post.

PE1 knows through some out-of-band mechanism (could be via BGP, Radius, manual configs, etc) that PE2, PE3 and PE4 are the egress nodes that belong to the same VPLS domain. PE1 now needs to establish an HSMP LSP (can be trivially extended to support a pseudowire) to PE2, PE3 and PE4. Figure shows 3 HSMP LSPs that will be required in this arrangement. The red HSMP LSP has PE1 as the root node and PE2, PE3 and PE4 as the leaf nodes. The green HSMP LSP has been initiated by PE2 and has PE2 as the root, and PE1, PE3 and PE4 as the leaf nodes. The blue HSMP LSP has been initiated by PE3 and has PE1, PE2 and PE4 as the leaf nodes. There is another HSMP LSP thats required – the one initiated by PE4 for the VPLS service to function. It has been omitted from the figure for the sake of clarity.

Thus all PE nodes in a VPLS service need to initiate an HSMP LSP (or a HSMP pseudowire) that terminates at the other PE routers.

In VPLS all BUM (broadcast, unknown unicast and multicast) traffic is flooded to all PE nodes. Its only the learnt traffic thats sent unicast by one PE to the other PE.

As explained earlier, a single copy sent by PE1 over an HSMP LSP will reach PE2, PE3 and PE4 (due to its P2MP component). Also the PE routers PE2, PE3 and PE4 can use this HSMP LSP (terminating at them) to send unicast traffic back to PE1.

Thus PE1 sends all BUM traffic on the HSMP LSP it initiates and all learnt unicast over the HSMP LSP that terminates on it.

Going back to figure 3, we can see that PE1 can use the red HSMP LSP to send all BUM traffic. This way it only sends one copy, and all the PE routers receive this packet. If PE1 wants to send learnt unicast traffic back to PE2, it uses the green HSMP LSP that terminates on it. PE1 can use this to send traffic back to PE2, which is the root node for this HSMP LSP. Similarly, PE1 uses the blue HSMP LSP whenever it wants to send learnt unicast traffic back to PE3.

Lets look at LSPs that PE2 uses. Whenever it wants to send BUM traffic, it uses the green HSMP LSP that it has originated. Any packet sent over that LSP is received by all the leaf nodes (which in this case happen to be PE1, PE3 and PE4). Like PE1, if it has to send learnt unicast traffic back to PE1, it uses the red HSMP LSP that was originated by PE1 and terminates at PE2.

Thus for a VPLS to fully function, all PE nodes must establish an HSMP LSP with all the other participating PE routers. It can use the optimized HSMP LSP that it originates for the BUM traffic and the HSMP LSP that other PEs originate for unicast communication.

The above table compares a VPLS service using HSMP LSPs with a regular VPLS service or Hierarchical VPLS (H-VPLS) service. Clearly, the former wins against the regular VPLS and H-VPLS on all counts. This may also perhaps be an answer to Juniper’s claim that H-VPLS is not scalable. Operators now need not implement H-VPLS; they can instead go in for VPLS services implemented using HSMP LSPs.

This post only briefly explains the idea behind HSMP LSPs. Its explained in detail here in this draft.