Remote-LFA Node Protection and Manageability
Individual Contributorpushpasis.ietf@gmail.comJuniper Networks, Inc.Electra, Exora Business ParkBangaloreKA560103Indiashraddha@juniper.netJuniper Networks, Inc.1194 N. Mathilda Ave.SunnyvaleCA94089UScbowers@juniper.netRtBrick, Inc.hannes@rtbrick.comOrangestephane.litkowski@orange.com
Routing
Routing Area Working GroupLFARemote-LFAIGPNode ProtectionThe loop-free alternates computed following the current
Remote-LFA specification guarantees only link-protection. The resulting
Remote-LFA nexthops (also called PQ-nodes), may not guarantee
node-protection for all destinations being protected by it.This document describes an extension to the Remote Loop-Free based
IP fast reroute mechanisms described in ,
that describes procedures for determining if a given PQ-node provides
node-protection for a specific destination or not. The document also
shows how the same procedure can be uitilized for collection of complete
characteristics for alternate paths. Knowledge about the characteristics
of all alternate path is precursory to apply operator defined policy
for eliminating paths not fitting constraints.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 RFC2119.The Remote-LFA
specification provides loop-free alternates that guarantee only
link-protection. The resulting Remote-LFA alternate nexthops (also
referred to as the PQ-nodes) may not provide node-protection for
all destinations covered by the same, in case of failure of the
primary nexthop node. Neither does the specification provide a means
to determine the same.Also, the
LFA Manageability document requires a computing router to
find all possible (including all possible Remote-LFA) alternate nexthops,
collect the complete set of path characteristics for each alternate path,
run an alternate-selection policy (configured by the operator) and find
the best alternate path. This will require the Remote-LFA implementation
to gather all the required path characteristics along each link on the
entire Remote-LFA alternate path.With current LFA and Remote-LFA implementations,
the forward SPF (and reverse SPF) is run with the computing router and its
immediate 1-hop routers as the roots. While that enables computation of
path attributes (e.g. SRLG, Admin-groups) for first alternate path segment
from the computing router to the PQ-node, there is no means for the computing
router to gather any path attributes for the path segment from the PQ-node to
destination. Consequently any policy-based selection of alternate paths
will consider only the path attributes from the computing router up until
the PQ-node. This document describes a procedure for determining node-protection
with Remote-LFA. The same procedure is also extended for collection
of a complete set of path attributes, enabling more accurate policy-based
selection for alternate paths obtained with Remote-LFA. Node-protection is required to provide protection of traffic on a
given forwarding node, against the failure of the first-hop node on the
primary forwarding path. Such protection becomes more critical in the
absence of mechanisms like non-stop-routing in the network. Certain
operators refrain from deploying non-stop-routing in their network,
due to the required complex state synchronization between redundant
control plane hardwares it requires, and the significant additional
performance complexities it hence introduces. In such cases node-protection
is essential to guarantee un-interrupted flow of traffic, even in the
case of an entire forwarding node going down. The following sections discuss the node-protection problem in the
context of Remote-LFA and propose a solution. To better illustrate the problem and the solution proposed in this
document the following topology diagram from the
Remote-LFA
draft is being re-used with slight modification.In the above topology, for all (non-ECMP) destinations reachable via
the S-E link there is no standard LFA alternate. As per the
Remote-LFA alternate
specifications node R2 being the only PQ-node for the S-E link provides
nexthop for all the above destinations.
below, shows all possible primary and Remote-LFA alternate paths for each
destination.DestinationPrimary PathPQ-nodeRemote-LFA Backup PathR3S->E->R3R2S=>N=>R1=>R2->R3ES->ER2S=>N=>R1=>R2->R3->ED1S->E->D1R2S=>N=>R1=>R2->R3->E->D1D2S->E->R3->D2R2S=>N=>R1=>R2->R3->D2A closer look at shows that, while the
PQ-node R2 provides link-protection for all the destinations, it does
not provide node-protection for destinations E and D1. In the event of
the node-failure on primary nexthop E, the alternate path from Remote-LFA
nexthop R2 to E and D1 also becomes unavailable. So for a Remote-LFA
nexthop to provide node-protection for a given destination, it is
mandatory that, the shortest path from the given PQ-node to the given
destination MUST NOT traverse the primary nexthop. In another extension of the topology in
let us consider an additional link between N and E with the same cost as the
other links. In the above topology, the S-E link is no more on any of the shortest
paths from N to R3, E and D1. Hence R3, E and D1 are also included in both
the Extended-P space and Q space of E (w.r.t S-E link).
below, shows all possible primary and R-LFA alternate paths via PQ-node R3, for
each destination reachable through the S-E link in the above topology. The
R-LFA alternate paths via PQ-node R2 remains same as in .DestinationPrimary PathPQ-nodeRemote-LFA Backup PathR3S->E->R3R3S=>N=>E=>R3ES->ER3S=>N=>E=>R3->ED1S->E->D1R3S=>N=>E=>R3->E->D1D2S->E->R3->D2R3S=>N=>E=>R3->D2Again a closer look at shows that, unlike
, where the single PQ-node R2 provided node-protection
for destinations R3 and D2, if we choose R3 as the R-LFA nexthop, it does not
provide node-protection for R3 and D2 anymore. If S chooses R3 as the R-LFA
nexthop, in the event of the node-failure on primary nexthop E, on the alternate
path from S to R-LFA nexthop R3, one of parallel ECMP path between N and R3
also becomes unavailable. So for a Remote-LFA nexthop to provide node-protection
for a given destination, it is also mandatory that, the shortest paths from S
to the chosen PQ-node MUST NOT traverse the primary nexthop node. This document adds and enhances the following definitions extending
the ones mentioned in Remote-LFA specification. The Remote-LFA specification
already defines this. The link-protecting extended P-space for a link S-E
being protected is the set of routers that are reachable from one or more
direct neighbors of S, except primary node E, without traversing the S-E
link on any of the shortest paths from the direct neighbor to the router.
This MUST exclude any direct neighbor for which there is at least one ECMP
path from the direct neighbor traversing the link(S-E) being protected. For a cost-based definition for Link-protecting Extended P-Space refer
to . The node-protecting extended P-space for a primary nexthop node E being
protected, is the set of routers that are reachable from one or more direct
neighbors of S, except primary node E, without traversing the node E. This
MUST exclude any direct neighbors for which there is at least one ECMP path
from the direct neighbor traversing the node E being protected. For a cost-based definition for Node-protecting Extended P-Space refer
to . The Remote-LFA draft
already defines this. The Q-space for a link S-E being protected is the
set of nodes that can reach primary node E, without traversing the S-E
link on any of the shortest paths from the node itself to primary nexthop E.
This MUST exclude any node for which there is at least one ECMP
path from the node to the primary nexthop E traversing the link(S-E)
being protected. For a cost-based definition for Q-Space refer to
. A node Y is in link-protecting PQ space w.r.t the link (S-E) being
protected, if and only if, Y is present in both link-protecting extended
P-space and the Q-space for the link being protected. A node Y is in candidate node-protecting PQ space w.r.t the node (E)
being protected, if and only if, Y is present in both node-protecting extended
P-space and the Q-space for the link being protected. Please note, that a node Y being in candidate node-protecting
PQ-space, does not guarantee that the R-LFA alternate path via the same, in
entirety, is unaffected in the event of a node failure of primary nexthop node
E. It only guarantees that the path segment from S to PQ-node Y is unaffected
by the same failure event. The PQ-nodes in the candidate node-protecting
PQ space may provide node protection for only a subset of destinations
that are reachable through the corresponding primary link. This section provides cost-based definitions for some of the terms introduced in
of this document. Please refer to for a formal
definition for Link-protecting Extended P-Space. A node Y is in link-protecting extended P-space w.r.t the link
(S-E) being protected, if and only if, there exists at least one direct
neighbor of S, Ni, other than primary nexthop E, that satisfies the
following condition. Please refer to for a formal
definition for Node-protecting Extended P-Space. A node Y is in node-protecting extended P-space w.r.t the node
E being protected, if and only if, there exists at least one direct neighbor
of S, Ni, other than primary nexthop E, that satisfies the following
condition. Please note, that a node Y satisfying the condition in
above only guarantees that the R-LFA alternate
path segment from S via direct neighbor Ni to the node Y is not affected
in the event of a node failure of E. It does not yet guarantee that the path
segment from node Y to the destination is also unaffected by the same
failure event. Please refer to for a formal definition
for Q-Space. A node Y is in Q-space w.r.t the link (S-E) being protected,
if and only if, the following condition is satisfied. The R-LFA alternate path through a given PQ-node to a given
destination is comprised of two path segments as follows.Path segment from the computing router to the
PQ-node (Remote-LFA alternate nexthop), andPath segment from the PQ-node to the destination
being protected. So to ensure a R-LFA alternate path for a given destination provides
node-protection we need to ensure that none of the above path segments are
affected in the event of failure of the primary nexthop node. Sections
and show
how this can be ensured. To choose a node-protecting R-LFA nexthop for a destination R3, router S
needs to consider a PQ-node from the candidate node-protecting PQ-space for
the primary nexthop E on shortest path from S to R3. As mentioned in
, to consider a PQ-node as candidate node-protecting
PQ-node, there must be at least one direct neighbor Ni of S, such that all
shortest paths from Ni to the PQ-node does not traverse primary nexthop
node E. Implementations SHOULD run the inequality in for all direct neighbors, other than primary nexthop
node E, to determine whether a node Y is a candidate node-protecting PQ-node.
All of the metrics needed by this inequality would have been already collected
from the forward SPFs rooted at each of direct neighbor S, computed as part of
standard LFA implementation. With reference to the
topology in , below
shows how the above condition can be used to determine the candidate node-protecting
PQ-space for S-E link (primary nexthop E).Candidate PQ-node (Y)Direct Nbr (Ni)D_opt (Ni,Y)D_opt (Ni,E)D_opt (E,Y)Condition MetR2N2 (N,R2)1 (N,E)2 (E,R2)YesR3N2 (N,R3)1 (N,E)1 (E,R3)NoAs seen in the above , R3 does not meet the
node-protecting extended-p-space inequality and so, while R2 is in candidate
node-protecting PQ space, R3 is not.Some SPF implementations may also produce a list of links and nodes traversed
on the shortest path(s) from a given root to others. In such implementations,
router S may have executed a forward SPF with each of its direct neighbors as
the SPF root, executed as part of the standard LFA
computations. So S may re-use the list of links and nodes collected from the same
SPF computations, to decide whether a node Y is a candidate node-protecting PQ-node or
not. A node Y shall be considered as a node-protecting PQ-node, if and only if,
there is at least one direct neighbor of S, other than the primary nexthop
E, for which, the primary nexthop node E does not exist on the list of nodes
traversed on any of the shortest paths from the direct neighbor to the
PQ-node. below is an illustration of the mechanism with
the topology in .Candidate PQ-nodeRepair Tunnel Path(Repairing router to PQ-node)Link-ProtectionNode-ProtectionR2S->N->R1->R2YesYesR2S->E->R3->R2NoNoR3S->N->E->R3YesNoAs seen in the above while R2 is candidate node-protecting
Remote-LFA nexthop for R3 and D2, it is not so for E and D1, since the
primary nexthop E is in the shortest path from R2 to E and D1. Once a computing router finds all the candidate node-protecting PQ-nodes for a
given directly attached primary link, it shall follow the procedure as proposed
in this section, to choose one or more node-protecting R-LFA paths, for
destinations reachable through the same primary link in the primary SPF graph. To find a node-protecting R-LFA path for a given destination, the computing
router needs to pick a subset of PQ-nodes from the candidate node-protecting PQ-space for
the corresponding primary nexthop, such that all the path(s) from the PQ-node(s)
to the given destination remain unaffected in the event of a node failure of
the primary nexthop node. To determine whether a given PQ-node belongs to such a
subset of PQ-nodes, the computing router MUST ensure that none of the primary nexthop
node are found on any of the shortest paths from the PQ-node to the given
destination. This document proposes an additional forward SPF computation for each of
the PQ-nodes, to discover all shortest paths from the PQ-nodes to the destination.
This will help determine, if a given primary nexthop node is on the shortest paths
from the PQ-node to the given destination or not. To determine if a given candidate
node-protecting PQ-node provides node-protecting alternate for a given destination,
or not, all the shortest paths from the PQ-node to the given destination has to be
inspected, to check if the primary nexthop node is found on any of these shortest paths.
To compute all the shortest paths from a candidate node-protecting PQ-node to one
(or more) destination, the computing router MUST run the forward SPF on the candidate
node-protecting PQ-node. Soon after running the forward SPF, the computer router
SHOULD run the inequality in below, once for
each destination. A PQ-node that does not qualify the condition for a given
destination, does not guarantee node-protection for the path segment from the
PQ-node to the specific destination.All of the above metric costs except D_opt(Y, D), can be obtained with
forward and reverse SPFs with E(the primary nexthop) as the root, run as
part of the regular LFA and Remote-LFA implementation. The Distance_opt(Y, D)
metric can only be determined by the additional forward SPF run with PQ-node
Y as the root. With reference to the topology in ,
below shows how the above condition can be used
to determine node-protection with node-protecting PQ-node R2.Destination (D)Primary-NH (E)D_opt (Y, D)D_opt (Y, E)D_opt (E, D)Condition MetR3E1 (R2,R3)2 (R2,E)1 (E,R3)YesEE2 (R2,E)2 (R2,E)0 (E,E)NoD1E3 (R2,D1)2 (R2,E)1 (E,D1)NoD2E2 (R2,D2)2 (R2,E)1 (E,D2)YesAs seen in the above example above, R2 does not meet the node-
protecting inequality for destination E, and D1. And so, once again,
while R2 is a node-protecting Remote-LFA nexthop for R3 and D2,
it is not so for E and D1.In SPF implementations that also produce a list of links and nodes
traversed on the shortest path(s) from a given root to others, the
inequality in above need not be
evaluated. Instead, to determine whether a PQ-node provides node-protection
for a given destination or not, the list of nodes computed from forward SPF run
on the PQ-node, for the given destination, SHOULD be inspected. In case
the list contains the primary nexthop node, the PQ-node does not
provide node-protection. Else, the PQ-node guarantees node-protecting
alternate for the given destination. Below is an illustration of the
mechanism with candidate node-protecting PQ-node R2 in the topology in
.DestinationShortest Path (Repairing router to PQ-node)Link-ProtectionNode-ProtectionR3R2->R3YesYesER2->R3->EYesNoD1R2->R3->E->D1YesNoD2R2->R3->D2YesYesAs seen in the above example while R2 is candidate node-protecting
R-LFA nexthop for R3 and D2, it is not so for E and D1, since the
primary nexthop E is in the shortest path from R2 to E and D1.The procedure described in this document helps no more than to
determine whether a given Remote-LFA alternate provides node-protection
for a given destination or not. It does not find out any new Remote-LFA
alternate nexthops, outside the ones already computed by standard
Remote-LFA procedure. However, in case of availability of more than one
PQ-node (Remote-LFA alternates) for a destination, and node-protection
is required for the given primary nexthop, this procedure will eliminate
the PQ-nodes that do not provide node-protection and choose only the ones
that does. In certain scenarios, when one or more destinations maybe reachable via
multiple ECMP (equal-cost-multi-path) nexthop nodes, and only link-protection
is required, there is no need to compute any alternate paths for such destinations.
In the event of failure of one of the nexthop links, the remaining primary
nexthops shall always provide link-protection. However, if node-protection is
required, the rest of the primary nexthops may not guarantee node-protection.
below shows one such example topology. In the above example topology, costs of all links are 1, except the following links:
Link: S-E1, Cost: 2Link: N-E2: Cost: 2Link: R1-R2: Cost: 2
In the above topology, on computing router S, destinations D1 and D2 are reachable
via two ECMP nexthop nodes E1 and E2. However the primary paths via nexthop node E2
also traverses via the nexthop node E1. So in the event of node failure of nexthop
node E1, both primary paths (via E1 and E2) becomes unavailable. Hence if
node-protection is desired for destinations D1 and D2, alternate paths that does
not traverse any of the primary nexthop nodes E1 and E2, need to be computed.
In the above topology the only alternate neighbor N does not provide such a LFA
alternate path. Hence one (or more) R-LFA node-protecting alternate paths for
destinations D1 and D2, needs to be computed. In the above topology, following are the link-protecting PQ-nodes.
Primary Nexthop: E1, Link-Protecting PQ-Node: { R2 } Primary Nexthop: E2, Link-Protecting PQ-Node: { R2 } To find one (or more) node-protecting R-LFA paths for destinations D1 and D2,
one (or more) node-protecting PQ-node(s) needs to be determined first. Inequalities
specified in and
can be evaluated to compute the node-protecting PQ-space for each of the nexthop nodes
E1 and E2, as shown in below. To select a PQ-node as node-protecting
PQ-node for a destination with multiple primary nexthop nodes, the PQ-node MUST satisfy
the inequality for all primary nexthop nodes. Any PQ-node which is NOT node-protecting
PQ-node for all the primary nexthop nodes, MUST NOT be chosen as the node-protecting
PQ-node for destination.Primary Nexthop (E)Candidate PQ-node (Y)Direct Nbr (Ni)D_opt (Ni,Y)D_opt (Ni,E)D_opt (E,Y)Condition MetE1R2N3 (N,R2)3 (N,E1)2 (E1,R2)YesE2R2N3 (N,R2)2 (N,E2)3 (E2,R2)Yes In SPF implementations that also produce a list of links and nodes
traversed on the shortest path(s) from a given root to others, the tunnel-repair
paths from the computing router to candidate PQ-node can be examined to
ensure that none of the primary nexthop nodes is traversed. PQ-nodes that provide
one (or more) Tunnel-repair paths(s) that does not traverse any of the primary nexthop
nodes, are to be considered as node-protecting PQ-nodes.
below shows the possible tunnel-repair paths to PQ-node R2.Primary-NH (E)PQ-Node (Y)Tunnel-Repair PathsExclude All Primary-NHE1, E2R2S==>N==>R1==>R2YesFrom and , in the above example, R2
being node-protecting PQ-node for both primary nexthops E1 and E2, should be chosen
as the node-protecting PQ-node for destinations D1 and D2 that are both reachable via
primary nexthop nodes E1 and E2. Next, to find a node-protecting R-LFA path from node-protecting PQ-node to
destinations D1 and D2, inequalities specified in
should be evaluated, to ensure if R2 provides a node-protecting R-LFA path for each of these
destinations, as shown below in . For a R-LFA path to qualify
as node-protecting R-LFA path for a destination with multiple ECMP primary nexthop
nodes, the R-LFA path from the PQ-node to the destination MUST satisfy the
inequality for all primary nexthop nodes.Destination (D)Primary-NH (E)PQ-Node (Y)D_opt (Y, D)D_opt (Y, E)D_opt (E, D)Condition MetD1E1R23 (R2,D1)2 (R2,E1)1 (E1,D1)NoD1E2R23 (R2,D1)3 (R2,E2)2 (E2,D1)YesD2E1R22 (R2,D2)2 (R2,E1)2 (E1,D2)YesD2E2R22 (R2,D2)2 (R2,E2)3 (E2,D2)Yes In SPF implementations that also produce a list of links and nodes
traversed on the shortest path(s) from a given root to others, the R-LFA
paths via node-protecting PQ-node to final destination can be examined to
ensure that none of the primary nexthop nodes is traversed. R-LFA path(s)
that does not traverse any of the primary nexthop nodes, guarantees
node-protection in the event of failure of any of the primary nexthop nodes.
below shows the possible R-LFA-paths for destinations D1
and D2 via the node-protecting PQ-node R2.Destination (D)Primary-NH (E)PQ-Node (Y)R-LFA PathsExclude All Primary-NHD1E1, E2R2S==>N==>R1==>R2 -->R3-->E1-->D1NoD2E1, E2R2S==>N==>R1==>R2 -->R3-->D2YesFrom and , in the
example above, the R-LFA path from R2 does not meet the node-protecting
inequality for destination D1, while it does meet the same inequality for
destination D2. And so, while R2 provides node-protecting R-LFA alternate
for D2, it fails to provide node-protection for destination D1. Finally,
while it is possible to get a node-protecting R-LFA path for D2, no such
node-protecting R-LFA path can be found for D1. In addition to the extra reverse SPF computations suggested by the
Remote-LFA draft (one reverse SPF for each of
the directly connected neighbors), this document proposes a forward SPF
computations for each PQ-node discovered in the network. Since the average
number of PQ-nodes found in any network is considerably more than the number
of direct neighbors of the computing router, the proposal of running one
forward SPF per PQ-node may add considerably to the overall SPF computation
time. To limit the computational overhead of the approach proposed, this
document proposes that implementations MUST choose a subset from the entire
set of PQ-nodes computed in the network, with a finite limit on the number
of PQ-nodes in the subset. Implementations MUST choose a default value
for this limit and may provide user with a configuration knob to override
the default limit. Implementations MUST also evaluate some default
preference criteria while considering a PQ-node in this subset. Finally,
implementations MAY also allow the user to override the default preference
criteria, by providing a policy configuration for the same. This document proposes that implementations SHOULD use a default
preference criteria for PQ-node selection which will put a score on each
PQ-node, proportional to the number of primary interfaces for which it
provides coverage, its distance from the computing router, and its
router-id (or system-id in case of IS-IS). PQ-nodes that cover more
primary interfaces SHOULD be preferred over PQ-nodes that cover fewer
primary interfaces. When two or more PQ-nodes cover the same number of
primary interfaces, PQ-nodes which are closer (based on metric) to the
computing router SHOULD be preferred over PQ-nodes farther away from it.
For PQ-nodes that cover the same number of primary interfaces and are
the same distance from the computing router, the PQ-node with smaller
router-id (or system-id in case of IS-IS) SHOULD be preferred. Once a subset of PQ-nodes is found, computing router shall run a forward
SPF on each of the PQ-nodes in the subset to continue with procedures proposed
in .With the regular Remote-LFA
functionality the computing router may compute more than one PQ-node
as usable Remote-LFA alternate nexthops. Additionally an alternate selection
policy may be configured to enable the network operator to choose one
of them as the most appropriate Remote-LFA alternate. For such policy-based
alternate selection to run, all the relevant path characteristics for each
the alternate paths (one through each of the PQ-nodes), needs to be
collected. As mentioned before in
the R-LFA alternate path through a given PQ-node to a given destination is
comprised of two path segments.The first path segment (i.e. from the computing router to the PQ-node)
can be calculated from the regular forward SPF done as part of standard and
remote LFA computations. However without the mechanism proposed in section
of this document, there is no way to
determine the path characteristics for the second path segment (i.e. from the
PQ-node to the destination). In the absence of the path characteristics for the
second path segment, two Remote-LFA alternate paths may be equally preferred
based on the first path segments characteristics only, although the
second path segment attributes may be different.The additional forward SPF computation proposed in
document shall also collect links, nodes and
path characteristics along the second path segment. This shall enable collection
of complete path characteristics for a given Remote-LFA alternate path to a
given destination. The complete alternate path characteristics shall then
facilitate more accurate alternate path selection while running the alternate
selection policy.As already specified in to limit the computational
overhead of the proposed approach, forward SPF computations MUST be run on a selected
subset from the entire set of PQ-nodes computed in the network, with a finite limit
on the number of PQ-nodes in the subset. The detailed suggestion on how to select
this subset is specified in the same section. While this limits the number of possible
alternate paths provided to the alternate-selection policy, this is needed to keep the
computational complexity within affordable limits. However if the alternate-selection
policy is very restrictive this may leave few destinations in the entire topology
without protection. Yet this limitation provides a necessary tradeoff between
extensive coverage and immense computational overhead.Many thanks to Bruno Decraene for providing his useful comments. We would also
like to thank Uma Chunduri for reviewing this document and providing valuable feedback.
Also, many thanks to Harish Raghuveer for his review and comments on the initial
versions of this document.N/A. - No protocol changes are proposed in this document.This document does not introduce any change in any of the
protocol specifications. It simply proposes to run an extra SPF rooted
on each PQ-node discovered in the whole network.