Multipath Traffic Engineering

Introduction Operators managing traffic within their networks have several tools, among them: Equal-cost Multipath (ECMP): balance traffic along multiple paths. This yields some resilience and some traffic management, as traffic can be load-balanced across multiple paths. To use ECMP effectively, one may have to adjust link metrics to allow multiple paths to have the same overall distance. Traffic Engineering (TE): state constraints for a path from an ingress router to an egress router, and let a path computation engine compute it. This gives much greater control over the nodes and links traversed, but is usually limited to finding a single path from ingress to egress . Multi-egress: allow traffic from an ingress router to a destination dst to use several egress routers, all of which have routes to that destination. dst may be an Internet prefix , a VPN prefix , an EVPN address , a VPLS site , or some other service destination. For BGP-signaled destinations, this requires that the BGP tie-breaking algorithm yield multiple results (rather than a single one), all of which become candidates for egress. describes requirements for MPLS-based TE, and thus is relevant to this memo. At the same time, the authors appear to believe that one can either have TE or multipathing, but not both. This is further emphasized by the notion of a Label Switched Path, which is used to implement MPLS-based TE. RSVP-TE (), the protocol designed to meet the requirements of , builds a single path from one ingress to one egress (for unicast traffic). In order to satisfy the constraints, TE often uses non-shortest paths. To do so without looplng packets, a tunnel is used. Such tunnels have to be signaled. RSVP-TE is a signaling protocol for MPLS-based tunnels. In this memo, we introduce a new tool: multipath TE (MPTE). This allows an operator to specify constraints for paths (as in TE), specify multiple egresses, and use multiple paths to each egress. Effectively, MPTE combines the advantages of the three tools above. The resulting set of paths from an ingress to egresses is a Directed Acyclic Graph (DAG), here called an MPTE DAG or MPTED. Finally, this memo allows the use of multiple types of tunnels. The main contribution of this memo is a protocol for signaling a (multipath) unicast tunnel across an MPTED.

Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Definition of Commonly Used Terms This section provides definitions for terms and abbreviations that have a specific meaning to the MPTE protocol and that are used throughout this memo.

constraints:: desired properties of paths between ingresses and egresses.
directed acyclic graph:: a directed graph that has no cycles.
directed graph:: a set of nodes and directed links. A network is represented by a directed graph.
egress:: an end node of an MPTE DAG.
ingress:: a starting node of an MPTE DAG.
link:: A (directed) edge between two nodes. A pair of nodes may have 0 or more links between them. A link between nodes u and v will be denoted by (u, v, i), where i is u's oif for the link. A link may have associated attributes, in particular, a metric.
metric:: a link attribute denoted by met(u, v, i), a positive number.
node:: a vertex of a graph. A node may have associated attributes.
outgoing interface:: a unique number (oif) assigned by a node for each outgoing link it has.
path length:: the sum of the metrics of the links that constitute path p, denoted by len(p)
shortest path:: a path between a pair of nodes u, v with minimum length. The set of shortest paths between u and v is a DAG, denoted by sp(u, v). The length of a shortest path from u to v is denoted by min(u, v)
slack:: a path p from u to v is acceptable with slack s if len(p) <= min(u, v) + s.
traffic trunk:: a unidirectional aggregate of traffic flows from an ingress to a set of egresses that is treated identically in the forwarding plane.

The following abbreviations are used in this memo:

CSPF:: constrained shortest path first. A modification to SPF to take into account path constraints.
DAG:: directed acyclic graph. The result of a multipath SPF or CSPF computation is a DAG.
ECMP:: equal-cost multipath.
FALB:: flow-aware load balancing
LB:: load balancing
LSP:: label-switched path
MC:: MPTED computer: the entity computing the MPTED, typically the ingress (if there is a single ingress) or a Path Computation Element
MPTE:: multipath TE with path constraints (including a slack) using nECMP paths from an ingress to one or more egresses.
MPTED:: an MPTE DAG resulting from CSPF-type computation on MPTE constraints.
MPTEP:: MPTE protocol: the protocol used to signal MPTEDs.
nECMP:: non-equal-cost multipath; generally qualified by "with slack s", meaning within slack s of the minimum path length.
oif:: outgoing interface index
PCE:: Path Computation Element
SPF:: shortest path first. Typically refers to Dijkstra's algorithm for computing shortest paths between a given pair of nodes, or pairwise between all nodes.
SRG:: shared risk group -- nodes and/or links that share "risk" (e.g., have common power, or use a common fiber conduit)
TE:: traffic engineering

Overview Consider :

Multipathing

ECMP (slack 0) from node 0 to node 5 There are 4 ECMP paths from node 0 to node 5: 0-2=3-5 (2 paths) 0-2-4-5 0-4-5 These 4 distinct paths all have length 300.

nECMP from node 0 to node 5 with slack 10 There are 7 nECMP paths with slack 10 to node 5: 0-2=3=5 (4 paths) 0-2-4-5 0-4-5 0-6-7-5 These 7 paths have lengths 300 or 310. Thus, allowing nECMP paths a slack of 10 has yielded 3 additional paths, which provide increased diversity and load balancing, and possibly decreased congestion.

Multipathing from node 0 to egresses {5, 8} If, for some traffic trunk that starts at node 0, nodes 5 and 8 are equally good as egresses, then one can compute an ECMPD from 0 to {5, 8}; this yields 4 paths to 5 and 6 paths to 8, for a total of 10 paths this traffic trunk can take. Similarly, a nECMP DAG to {5, 8} with slack 10 has 15 paths, whereas one with slack 5 has the same 11 paths as with slack 0.

MPTED from ingresses {0, 1} to egresses {5, 8} If traffic from node 0 to nodes {5, 8} and from node 1 to nodes {5, 8} have common characteristics, it may make sense to compute a single DAG from {0, 1} to {5, 8}. Doing so allows the operator to view this entire DAG as one logical entity; a nice side benifit is reduced control and data plane state due to state sharing.

Load balancing Nodes in a netword have a Forwarding Information Base (FIB). A FIB maps a packet's destination address da to one or more "next hops". When a packet with address da arrives at n, n sends the packet to one of the next hops. n typically will distribute packets in a given ratio among the next hops. This is load balancing. The main goal of ECMP/nECMP is to supply as many nodes as possible in the MPTED with multiple next hops on which to forward the traffic trunk. At such nodes, traffic belonging to the trunk can be distributed among the next hops instead of going to a single next hop. This has the potential to reduce congestion and provide better utilization of available links.

Flow-aware load balancing When load balancing packets from a traffic trunk, it is often required that packets from a given flow be sent to the same next hop. This improves the probability of in-order delivery of packets in that flow, which is important for certain types of traffic. This is called flow-aware load balancing (FALB). The most common flow in IP traffic is defined by a 5-tuple consisting of the source IP address, the destination IP address, the protocol, the source port and the destination port. A 16- or 20-bit hash of this 5-tuple is called the packet's entropy. There are two common ways to achieve FALB of IP traffic. One is to do a "deepish" packet inspection (dPI), find the relevant 5-tuple, and use that to compute the packet's entropy. The entropy is then used to ensure that packets in the flow are sent to the same next hop. This memo suggests sending TE traffic over a tunnel (see {tunnels}); this makes the identification of IP flows expensive and error-prone. Another way of accomplishing this is to insert the entropy in the tunnel header. Many of the tunnels suggested in this memo have such a field. The ingress is in a good position to identify flows, and, when encapsulating the packet into the tunnel, can insert the entropy in the header. The heavy lifting of identifying flows is thus placed on the ingress. Transit nodes can simply use the entropy field to correctly map packets in a flow to the same next hop, thus ensuring FALB.

Per-packet load balancing FALB is often required and is a good default behavior, especially as end applications may be expecting packets in a flow to be delivered in order. However, FALB has the issue that it attempts (statistically) to place roughly the number of flows in the given ratio on the outgoing links; that may not place traffic in the same ratio, as flows need not carry the same traffic. In some cases (typically when configured to), one can do per-packet load balancing (PPLB), meaning that load balancing is no longer flow aware. This can be done when the end applications do not require packets in a flow to be in order, or if some (bookended) devices outside the network put the packets back in order before delivering them to the applications (typically by addind a sequence number). When feasible, PPLB gives much better load distribution, and is currently the subject of investigation, implementation and standardization. One can achieve this by configuring each router in the DAG to do PPLB for the traffic trunks in the DAG, or more simply by the ingress router assigning entropy at random to the traffic it places in the DAG. The latter approach keeps the decision of which DAGs (and corresponding traffic trunks) should be flow-aware and which not at the ingress; all other nodes simply do what the entropy fields tells them to do.

Constraints Constraints are an intent-based specification of acceptable paths that a traffic trunk may take from ingress to egress(es). Constraints are thus an abstract way to control the resources that a particular traffic trunk uses. One way to do this is to add "resource class attributes" or "colors" to links, and then specify "include" and "exclude" sets. An include set means that all links that a path traverses must contain at least one element of the include set. An exclude set means that no link in the path can contain any color from the exclude set. Another way is to specify a (maximum) bandwidth that a traffic trunk can carry. This means that all links in the path must have that much available capacity. Packets exceending the bandwidth can forwarded normally, marked as droppable, or dropped. Let's add some simple constraints to our DAG. We associate the color red to one of the links from B to C, and to the shorter of the links from F to G. Then, we constrain the paths to "exclude red", meaning avoid links with color red. This yields the following: ECMP from node 0 to node 5 with constraints "include red or blue" yields a single path.

Protection One very useful aspect of TE is the ability to specify that a path must be link- or node- or shared-risk-disjoint from another path. That means that the two paths do not have links or nodes or "shared risk groups". Additionally, one can build protection paths for an existing path to protect against link or node failures . This is especially important as TE currently takes a single path through the network, meaning that a link or node failure will result in dropped traffic until the TE path is restored. While not quite as crucial in the case of an MPTED, since ideally, there will be multiple nexthops at each node, there will be cases where a node has a single next hop, or all next hops share a common failure mode. Identifying these cases and building protection paths for such nodes will be described in a future version of this memo.

Tunnels The shortest path first algorithm is an easy-to-implement and very efficient algorithm whereby all routers in a network can agree on the path that a packet to a particular destination should take. That means, if all routers are agreed (roughly) on the topology and metrics of the network, they will forward packets in a loop-free manner to all destinations -- without the need for signaling or tunnels. However, an MPTED will not take the same paths -- some paths may be rejected as they don't conform to the constraints, and others may be used even though they are not shortest paths. Thus, to route packets in a traffic trunk over a computed MPTED, a tunnel is typically used. This tunnel will have to be signaled to the MPTED nodes. The tunnel may be MPLS- or IP-based. A few things are important about tunnels: whether they carry an entropy field (EF), whether they have a "discriminator" (D) that allows multiple tunnels between an ingress-egress pair, whether they allow multiple egresses (ME), and whether they allow multiple ingresses (MI). These will be discussed in the description of the tunnels below. In the memo, we consider the following tunnel types: IP-in-IP: encapsulation allows the creation of an "outer" IP header to carry a payload packet (which is typically an IP payload). The outer IP header's protocol field indicates the "protocol" of the inner payload packet. The outer header of IP-in-IP tunnel doesn't contain an EF; transit nodes can either spray packets across outgoing next hops, attempt to do dPI, or use the same next hop for all packets. To accommodate ME, the egresses have to have the same (anycast) IP address which would be used as the destination IP of the tunnel. MI is not possible. GRE: Generic Routing Encapsulation. We include in this definition and with the Key Present (bit 2) set to 0. This is similar to IP-in-IP; however, the payload is not required to be IP. There is no EF in the header. D, ME and MI same as for IP-in-IP. GRE-E: GRE with Key Present; the Key value is the EF. D, ME and MI same as for IP-in-IP. GRE6: GRE with IPv6 addresses. The entropy is carried in the Flow Label field of the IPv6 header. D, ME and MI same as for IP-in-IP. G-in-U: GRE-in-UDP . The UDP source port is the EF; the GRE Key, if present, can be ignored from a load balancing point of view. D, ME and MI as in IP-in-IP. MPLS-in-UDP . The UDP source port is the EF; D, ME and MI as in IP-in-IP. SigLab (signaled label switching). The labels to be used are signaled. Signaling proceeds from egress(es) to ingress(es). An entropy label can be used as the EF. At each node, a different label is used for each MPTED; this is the discriminator. ME and MI are both allowed. StatLab (static label). A single statically-assigned label defines the tunnel throughout the MPTED. Here, a block of MPLS labels is given to a label allocator; these labels MUST NOT be allocated by any node in the network. EF, D, ME and MI are as for SigLab. The MPTED computer (MC) must interact with the allocator when creating or deleting an MPTED.

Operation The starting point in building an MPTE DAG is to define the properties of a traffic trunk from ingress to egress. Examples include "BGP destinations with community xyz" or "gold class traffic belonging to VPN foo". Next, define a set of constraints that capture the types of paths permissible for this traffic trunk. These include a metric to minimize (perhaps with slack); this could capture delay or fiber length, link colors, shared risk groups (SRGs) and bandwidth. The desired outcome is an MPTED into which the traffic trunk can be mapped. An MPTED is specified by defining: a (non-empty) set of ingresses a (non-empty) set of egresses the metric to use and the slack path constraints whether or not the MPTED is "strict". An MPTED is strict if all paths from all ingresses to all egresses are within slack of the shortest path. An MPTED is loose if all paths from a given ingress I to a given egress E are within slack of each other, but paths from I to a different egress F may not be within slack of the paths to I. Computation (possibly using a variant of CSPF) of an MPTED is done by the MC, which is either an ingress or a PCE . (This memo does not specify such an algorithm.) Signaling primarily occurs between the MC and each junction node. Auxiliary signaling may occur between a junction node and its phops.

MPTED In this memo, a node is identified by its (16-octet) IPv6 loopback address. A link from node u to node v is identified by u's loopback address and its (4-octet) outgoing interface index (oif), a unique identifier for the link allocated by u. oifs are usually exchanged in the TE extensions of an IGP. (A link also has a (4-octet) incoming interface index, the iif. For neighbors u and v, the correlation between u's oif and v's iif is typically done by the IGP. iifs are not used in this memo.) For now, this memo only deals with point-to-point links; a future revision will describe the use of multi-access links. An MPTED is identified by a unique (4-octet) ID (the MID) assigned to the MPTED by the MC. As an MPTED can change over its lifetime, it is assigned a version number starting at 0 and incremented every time the MPTED is recomputed. Thus, a full MPTED ID (the FID) consists of <MC, MID, version>. An MPTED consists two or more "junction nodes". A junction node can have one of five types: a pure ingress node has zero incoming links and one or more outgoing links in the MPTED. Traffic routed on a MPTED enters at the ingress. a pure egress node has one or more incoming links and zero outgoing links in the MPTED. Traffic routed on a MPTED leaves at an egress. a transit ingress node where traffic can either enter the MPTED or arrive from another ingress node to continue on in the MPTED. a transit egress node where traffic can either exit the MPTED or go on to another egress node. a "regular" junction node has one or more incoming links and one or more outgoing links. Traffic does not enter or leave at such a node: it comes from a phop and goes to an nhop. A junction node v consists of v, its previous hops (phops) and its next hops (nhops). A phop is specified by an incoming link of v: (u, v, oif1); an nhop by an outgoing link of v: (v, w, oif2). Note that, since links are point-to-point, it is sufficient to specify (u, oif1) ((v, oif2)) for a phop (nhop, respectively). The nodes u (and w) are loosely referred to as a phop (and nhop) of v, although strictly speaking the link should be included. A pure ingress has no phops and a pure egress has no nhops. The MPTED is broken down into a set of junction nodes. A junction node v is specified by: bandwidth (coming in to and going out of v) a list of phops of v a list of nhops of v, with corresponding load balancing splits

Signaling overview The MC signals the creation or update of an MPTED by sending to each junction node v a JUNCTION message consisting of: the MPTED ID the junction node specification the tunnel type some flags After v parses this specification, for all tunnel types other than SigLab, it installs FIB state for the junction. For tunnel type SigLab, v allocates an incoming MPLS label L_u for each phop u, and sends a LABEL message to u containing: the MPTED ID the phop (u's loopback + u's oif for the link) the allocated label L_u u records label L_u as part of its own junction state. When v receives a LABEL message from all its nhops, it installs swap state in its LFIB.

Forwarding state <TBD>

Protocol MPTEP, the protocol used to create an MPTED, runs over TCP, and is loosely modeled on BGP . The following TCP sessions are needed: between any ingress acting as MC and all potential junction nodes; between the PCE and all potential junction nodes; if tunnel type SigLab is used, between a junction node and all its immediate neighbors. Thus, there will be a full mesh of TCP sessions between all pairs of potential junction nodes. For networks with several hundreds or thousands of nodes, see for an alternative solution.

Message IDs Every semantically significant message (SSM) (i.e., one that causes state to be created in a receiver) has a (4-octet) message ID (msgID). msgID starts from 1 and counts up in a session. The last processed and stored message ID is sent in a hello. This tells the sender of the SSM that the receiver of the SSM (sender of the hello) has finished processing the SSM. See .

Messages An MPTEP message consists of a fixed-length message header (including a message type) followed by a variable length body that depends on the type. There are two types of message headers, MSGHDR and REFHDR. MSGHDR MUST NOT be used for messages to or from a reflector. REFHDR MUST be used for all messages to or from a reflector.

MSGHDR A "normal" MPTEP message header has the following format:

Type:: message type (2 octets)
Length:: total length of the message (including header) in octets (2 octets)

REFHDR An MPTEP message to or from an MPTEP reflector uses the following header:

Message types

OPEN

Version:: 0 (MUST match between endpoints)
Capabilities:: bit vector of sender's capabilities
0x1:: node is capable of Graceful Restart (see )
Rest:: SHOULD be zero on sending and ignored on receipt
Hello Time:: time in seconds between hellos. If a hello is not received in time, it is deemed to be missed. If three consecutive hellos are missed, the session is torn down.
Keep Time:: time that control plane and forwarding plane state received from neighbor is kept after session teardown.
Sender Identifier, Receiver Identifier:: IPv6 loopback addresses of the two endpoints
Supported Tunnel Types:: bit vector of tunnel types that the sender can install. If the receiver is an MC, it MUST NOT send an MPTED with a tunnel type that the sender does not implement.
Opt Param Len, Optional Parameters:: none defined yet

HELLO

Indication that the sender is alive and functioning; also, that the sender has processed and safely stored state related to messages up to and including the enclosed msgID; the receiver can throw away signaling state for messages with a lower msgID.

JUNCTION A JUNCTION message has the following format:

The Tunnel Information field is to identify the type of tunnel to use for the MPTED. For example, for an MPLS tunnel with a statically assigned label, the Tunnel Information is the label. For IP-based tunnels, the Tunnel Information is the source and destination IP addresses (plus possibly other information). Details TBD. The fields marked (?) may not be required in a Junction message; TBD.

Tunnel Flags (1 octet)

0x1:: Junction is an ingress
0x2:: Junction is an egress

(Pure vs. transit ingresses/egresses are distinguished by the number of phops/nhops.)

Rest:: Reserved (MUST be sent as 0 and ignored on receipt)

Junction bandwidth bandwidth incoming to the junction in Megabits per second (Mbps) as a 4 octet non-negative integer

nhop share 2-octet share of the outgoing bandwidth. A Junction should attempt to send a ratio of (share n)/(sum (share i)) of the incoming bandwidth to nhop #n.

LABEL A LABEL message MUST only be used for MPTEDs of type SigLab. A LABEL message is sent from an egress junction node to each of its phops. Any other junction node MUST only send a LABEL message when it has received a LABEL message from all of its nhops (cf "Ordered Label Distribution Control" ). A pure ingress node never sends a LABEL message as it has no phops.

NOTIFICATION <TBD>

MPTEP Reflector Instead of establishing a full mesh of MPTEP connections among the nodes participating in establishing MPTE DAGs, one could instead have a small set of designated MPTEP Reflectors with whom all MPTEP nodes establish connections. An MPTEP Reflector passes on an MPTEP message from the sender to the (single) ultimate receiver. In this, an MPTEP's function is different from that of a BGP Reflector: an MPTEP Reflector sends a received message to exactly one destination node. The goal of having MPTEP Reflectors is simply to reduce the number of MPTEP sessions that a node (typically, a router) has. In a network of (say) 500 nodes and (say) 3 Reflectors, each of these 500 nodes would only need 3 sessions with the Reflectors. The Reflectors themselves would need 500 sessions with the router nodes, plus 2 sessions among themselves.

Graceful Restart A node N is capable of Graceful Restart if a) it can maintain control plane state across restarts; and b) it can maintain forwarding state across restarts. If N is capable of Graceful Restart, an MPTE DAG going through N can continue functioning while N restarts. While N is restarting, new JUNCTION/LABEL messages will be dropped or ignored; new MPTE DAGs passing through N will not be established. Once restart is complete, N will send an OPEN message and re-establish connections will all its peers (or all the MPTEP Reflectors). Thereafter, N can participate in new DAGs passing through it by processing received JUNCTION messages. More details will be described in a future version.

IANA Considerations TBD

Security Considerations TBD