high-level: MPLS is a way to speed up forwarding for IP packets within a particular domain. it does this by labelling packets on ingress (with a fixed-length header) and then forwarding them according to their label until they exit the MPLS domain. there are various reasons to do this, including:
- speeding up IP packet forwarding
- moving routing logic into the data plane
- implementing differentiated service / policy routing
- fast failure restoration
- encoding “alternate routes”
some of them these are design goals of MPLS, and some of them are somewhat incidental.
how MPLS works
MPLS is a shim between the link and network layers. it can work with any protocol above or below it, hence multi-protocol.
MPLS headers
it has its own 32-bit header that it inserts between the network and link headers. this header has:
- 20 bits for the label
- 3 experimental bits
- 1 bit for the stack
- 8 bits for the TTL (to avoid routing loops)
the label defines a Forwarding Equivalence Class.
Forwarding Equivalence Class
Definition: a subset of packets treated the same way by a router.
in the context of MPLS, it is specifically the subset of packets that are treated the same way by a LSR.
Link to original
MPLS operation
- at ingress, compute the MPLS header and insert it into the packet
- at each LSR:
- use the label to index into a forwarding table that specifies:
- the next hop, and
- the new label to attach to the packet
- replace the old label with the new label and forward the packet
- use the label to index into a forwarding table that specifies:
- at egress, get rid of the MPLS header and forward the packet based on IP/whatever else
the above description doesn’t specify how to decide or communicate the forwarding information. there are two broad ways: hop-by-hop routes that query the dynamic routing protocol (ex Open Shortest Path First), or explicit routes where the whole path is set by the LSR.
how is forwarding info decided? probably by the label distribution protocol? i don't know
keywords: frame relay — trade delay for optimal path
downstream benefits and uses
MPLS labels can be used with differentiated service, letting you provide QoS. this is because you can configure LSPs between each ingress/egress pair, so by either using the exp bits or by creating an LSP for each traffic class you can implement differentiated service in that bit of the network. this decouples the number of traffic-class-paths from the number of flows by using MPLS’ labels, so it scales better. it also still interoperates with the DS field-based architecture (what is this?).
MPLS also enables fast failure restoration. specifically, we can get efficient local protection, where you keep around a backup path (and reserve resources, according to your SLA) for every possible link/node failure. the router that detects the failure will switch traffic to the backup router. this restores faster than end-to-end protection, since you can restore in the middle of the path rather than restarting from the source router.
rough notes
- instead of using ip masks for forwarding (longest prefix match), you circumvent the problem entirely
- try to get at source routing via mpls
- concept: use small, fixed-length field, so can decide forwarding just by indexing (fast in hardware)
- adding a thin layer between link and network layer
- “shim”
- this is where “multi-protocol” comes in — you don’t necessarily have to use IP on top or something?
- this basically lets you do some routing in the data plane
- MPLS headers have 32 bits
- 20 bits for label
- 3 bits for “stuff” (experimental)
- stack..? (1 bit)
- TTL (since you still want to avoid loops)
- build a label-indexed forwarding table
- i.e. label switching router
- forwarding a packet involves more complexity than just “straight on through”
- often you want some extra information? in the forwarding? some tying between network and link layers
- “differentiated service” or something
- and this is complex and annoying to do on every hop
- and so MPLS lets you sort sets of packets that are all treated the same by an label-switched router (forwarding equivalence classes)
- i.e. when you enter the MPLS network (ingress), you decide all this stuff once, include that info in the label, and then you can do all your policy stuff fast in hardware with simple label lookup
ok so how does this work
- packet enters the mpls domain
- you stick in the shim (i.e. encode the packet’s forwarding equivalence class into the label)
- these labels have local significance (cause 20 bits is too little to be unique)
- so the switching operation is now:
- (interface in, input label, interface out, label out) entries in table
- forward packet to interface out, and map label in → out
- something something label distribution?
- what does this do i have no idea
- ok now how do you decide these routes within the MPLS network
- can query the routing protocol at each hop: tell me the next hop
- or you can do explicit routing
- i.e. at ingress, the control protocol sets the route that should be used
- so you can deal with “other information” (policy, differentiated service, etc etc)
- protocols: CR-LDP, RSVP-TE
ok so why isnt this used everywhere
- what happens when things break?
- state is installed somehow, these forwarding tables are state, what happens if the state changes
- switch breaks, how do you reroute when you do source routing?
- this is kind of fine when the paths are dynamically installed by routing protocols, but harder when you do end-system-installed routes
- IP restoration you need to wait for the routing protocols to update with protocols
- MPLS also gives you fast failure restoration
- you also set up a backup path in advance
- link- and node-disjoint with the main path
- but this needs fast notification of failure
- aside on failure detection * its nontrivial to do this * but actually you get this implicitly in the PHY
- the amount of resources you dedicate to do this restoration is defined by the SLA
- right so this needs “local protection” which means you need to distinguish backup paths
- so this is what the stack is for