]>
Algorithms for computing Maximally Redundant Trees for IP/LDP Fast-RerouteJuniper Networks10 Technology Park DriveWestfordMA01886USAakatlas@juniper.netEricssonKonyves Kalman krt 11BudapestHungary1097Gabor.Sandor.Enyedi@ericsson.comEricssonKonyves Kalman krt 11BudapestHungary1097Andras.Csaszar@ericsson.comUniversity of Arizona1230 E Speedway Blvd.TucsonUSA85721AZabishek@ece.arizona.edu
Routing
Routing Area Working GroupA complete solution for IP and LDP Fast-Reroute using Maximally
Redundant Trees is presented in
[I-D.ietf-rtgwg-mrt-frr-architecture]. This document describes
an algorithm that can be used to compute the necessary Maximally
Redundant Trees and the associated next-hops.MRT Fast-Reroute requires that packets can be forwarded not only
on the shortest-path tree, but also on two Maximally Redundant Trees
(MRTs), referred to as the Blue MRT and the Red MRT. A router which
experiences a local failure must also have pre-determined which
alternate to use. This document describes how to compute these
three things and the algorithm design decisions and rationale. The
algorithms are based on those presented in and expanded in .Just as packets routed on a hop-by-hop basis require that each
router compute a shortest-path tree which is consistent, it is
necessary for each router to compute the Blue MRT and Red MRT in a
consistent fashion. This is the motivation for the detail in this
document.As now, a router's FIB will contain primary next-hops for the
current shortest-path tree for forwarding traffic. In addition, a
router's FIB will contain primary next-hops for the Blue MRT for
forwarding received traffic on the Blue MRT and primary next-hops
for the Red MRT for forwarding received traffic on the Red MRT.What alternate next-hops a point-of-local-repair (PLR) selects
need not be consistent - but loops must be prevented. To reduce
congestion, it is possible for multiple alternate next-hops to be
selected; in the context of MRT alternates, each of those alternate
next-hops would be equal-cost paths.This document provides an algorithm for selecting an appropriate
MRT alternate for consideration. Other alternates, e.g. LFAs that
are downstream paths, may be prefered when available and that
decision-making is not captured in this document.Algorithms for computing MRTs can handle arbitrary network
topologies where the whole network graph is not 2-connected, as in
, as well as the
easier case where the network graph is 2-connected (). Each MRT is a spanning
tree. The pair of MRTs provide two paths from every node X to the
root of the MRTs. Those paths share the minimum number of nodes and
the minimum number of links. Each such shared node is a cut-vertex.
Any shared links are cut-links. A pair of trees where the
path from any node X to the root R on the first tree is
node-disjoint with the path from the same node X to the root
along the second tree. These can be computed in 2-connected
graphs.A pair of trees
where the path from any node X to the root R along the first tree
and the path from the same node X to the root along the second
tree share the minimum number of nodes and the minimum number of
links. Each such shared node is a cut-vertex. Any shared links
are cut-links. Any RT is an MRT but many MRTs are not RTs.A graph that reflects the network
topology where all links connect exactly two nodes and broadcast
links have been transformed into the standard pseudo-node
representation.A vertex whose removal partitions the
network.A link whose removal partitions the
network. A cut-link by definition must be connected between two
cut-vertices. If there are multiple parallel links, then they
are referred to as cut-links in this document if removing the set
of parallel links would partition the network. A graph that has no cut-vertices.
This is a graph that requires two nodes to be removed before the
network is partitioned.A tree containing links that
connects all nodes in the network graph.In the context of a spanning tree
computed via a depth-first search, a back-edge is a link that
connects a descendant of a node x with an ancestor of x.A maximal set of nodes that
are 2-connected. In a network graph with at least one
cut-vertex, there will be multiple 2-connected clusters.Either a 2-connected cluster, a cut-edge,
or an isolated vertex.Directed Acyclic Graph - a digraph containing
no directed cycle.Almost Directed Acyclic Graph - a digraph
that can be transformed into a DAG whith removing a single node
(the root node).Generalized ADAG - a digraph, which has only
ADAGs as all of its blocks. The root of such a block is the node
closest to the global root (e.g. with uniform link costs).Depth-First SearchA node n is a DFS ancestor of x if n
is on the DFS-tree path from the DFS root to x.A node n is a DFS descendant of x if x
is on the DFS-tree path from the DFS root to n.A path along not-yet-included-in-the-GADAG
nodes that starts at a node that is already-included-in-the-GADAG
and that ends at a node that is already-included-in-the-GADAG.
The starting and ending nodes may be the same node if it is a
cut-vertex.Indicates the
relationship between X and Y in a partial order, such as found in
a GADAG. X >> Y means that X is higher in the partial
order than Y. Y << X means that Y is lower in the partial
order than X. Indicates the relationship
between X and Y in the total order, such as found via a
topological sort. X > Y means that X is higher in the total
order than Y. Y < X means that Y is lower in the total order than X.A node added to the network graph to
represent a multi-homed prefix or routers outside the local
MRT-fast-reroute-supporting island of routers. The key property
of proxy-nodes is that traffic cannot transit them.There are five key concepts that are critical for understanding the
algorithms for computing MRTs. The first is the idea of partially
ordering the nodes in a network graph with regard to each other and to
the GADAG root. The second is the idea of finding an ear of nodes and
adding them in the correct direction. The third is the idea of a
Low-Point value and how it can be used to identify cut-vertices and to
find a second path towards the root. The fourth is the idea that a
non-2-connected graph is made up of blocks, where a block is a
2-connected cluster, a cut-edge or an isolated node. The fifth is the
idea of a local-root for each node; this is used to compute ADAGs in
each block.Given any two nodes X and Y in a graph, a particular total order
means that either X < Y or X > Y in that total order. An
example would be a graph where the nodes are ranked based upon their
IP loopback addresses. In a partial order, there may be some nodes
for which it can't be determined whether X << Y or X >> Y.
A partial order can be captured in a directed graph, as shown in . In a graphical representation, a link
directed from X to Y indicates that X is a neighbor of Y in the
network graph and X << Y.To compute MRTs, it is very useful to have the root of the MRTs be
at the very bottom and the very top of the partial ordering. This
means that from any node X, one can pick nodes higher in the order
until the root is reached. Similarly, from any node X, one can pick
nodes lower in the order until the root is reached. For instance, in
, from G the higher nodes picked can be
traced by following the directed links and are H, D, E and R.
Similarly, from G the lower nodes picked can be traced by reversing
the directed links and are F, B, A, and R. A graph that represents
this modified partial order is no longer a DAG; it is termed an Almost
DAG (ADAG) because if the links directed to the root were removed, it
would be a DAG.Most importantly, if a node Y >> X, then Y can only appear on
the increasing path from X to the root and never on the decreasing
path. Similarly, if a node Z << X, then Z can only appear on
the decreasing path from X to the root and never on the inceasing
path.Additionally, when following the increasing paths, it is possible
to pick multiple higher nodes and still have the certainty that those
paths will be disjoint from the decreasing paths. E.g. in the
previous example node B has multiple possibilities to forward packets
along an increasing path: it can either forward packets to C or F.For simplicity, the basic idea of creating a GADAG by adding ears
is described assuming that the network graph is a single 2-connected
cluster so that an ADAG is sufficient. Generalizing to multiple
blocks is done by considering the block-roots instead of the GADAG
root - and the actual algorithms given in and .In order to understand the basic idea of finding an ADAG, first
suppose that we have already a partial ADAG, which doesn't contain all
the nodes in the block yet, and we want to extend it to cover all the
nodes. Suppose that we find a path from a node X to Y such that X and
Y are already contained by our partial ADAG, but all the remaining
nodes along the path are not added to the ADAG yet. We refer to such a
path as an ear.Recall that our ADAG is closely related to a partial order, more
precisely, if we remove root R, the remaining DAG describes a partial
order of the nodes. If we suppose that neither X nor Y is the root, we
may be able to compare them. If one of them is definitely lesser with
respect to our partial order (say X<<Y), we can add the new path
to the ADAG in a direction from X to Y. As an example consider .In this partial ADAG, node C is not yet included. However, we can
find path B-C-D, where both endpoints are contained by this partial
ADAG (we say those nodes are *ready* in the sequel), and the remaining
node (node C) is not contained yet. If we remove R, the remaining DAG
defines a partial order, and with respect to this partial order we can
say that B<<D, so we can add the path to the ADAG in the
direction from B to D (arcs B->C and C->D are added). If B were
strictly greater than D, we would add the same path in reverse
direction.If in the partial order where an ear's two ends are X and Y, X
<< Y, then there must already be a directed path from X to Y
already in the ADAG. The ear must be added in a direction such that
it doesn't create a cycle; therefore the ear must go from X to Y.In the case, when X and Y are not ordered with each other, we can
select either direction for the ear. We have no restriction since
neither of the directions can result in a cycle. In the corner case
when one of the endpoints of an ear, say X, is the root (recall that
the two endpoints must be different), we could use both directions
again for the ear because the root can be considered both as smaller
and as greater than Y. However, we strictly pick that direction in
which the root is lower than Y. The logic for this decision is
explained in A partial ADAG is started by finding a cycle from the root R back
to itself. This can be done by selecting a non-ready neighbor N of R
and then finding a path from N to R that doesn't use any links between
R and N. The direction of the cycle can be assigned either way since
it is starting the ordering.Once a partial ADAG is already present, we can always add ears to it:
just select a non-ready neighbor N of a ready node Q, such that Q is
not the root, find a path from N to the root in the graph with Q
removed. This path is an ear where the first node of the ear is Q, the
next is N, then the path until the first ready node the path reached
(that second ready node is the other endpoint of the path). Since the
graph is 2-connected, there must be a path from N to R without Q.It is always possible to select a non-ready neighbor N of a ready
node Q so that Q is not the root R. Because the network is
2-connected, N must be connected to two different nodes and only one
can be R. Because the initial cycle has already been added to the
ADAG, there are ready nodes that are not R. Since the graph is
2-connected, while there are non-ready nodes, there must be a
non-ready neighbor N of a ready node that is not R.Algorithm merely requires that a
cycle or ear be selected without specifying how. Regardless of the
way of selecting the path, we will get an ADAG. The method used for
finding and selecting the ears is important; shorter ears result in
shorter paths along the MRTs. There are three options being considered.
The Low-Point Inheritance option is described in . The SPF-based option is described in
and the hybrid option is described in
.As an example, consider
again. First, we select the shortest cycle containing R, which can be
R-A-B-F-D-E (uniform link costs were assumed), so we get to the
situation depicted in (b). Finally, we
find a node next to a ready node; that must be node C and assume we
reached it from ready node B. We search a path from C to R without B
in the original graph. The first ready node along this is node D, so
the open ear is B-C-D. Since B<<D, we add arc B->C and
C->D to the ADAG. Since all the nodes are ready, we stop at this
point.A basic way of computing a spanning tree on a network graph is to
run a depth-first-search, such as given in . This tree has the important property that if
there is a link (x, n), then either n is a DFS ancestor of x or n is a
DFS descendant of x. In other words, either n is on the path from the
root to x or x is on the path from the root to n.Given a node x, one can compute the minimal DFS number of the
neighbours of x, i.e. min( D(w) if (x,w) is a link). This gives the
highest attachment point neighbouring x. What is interesting, though,
is what is the highest attachment point from x and x's descendants.
This is what is determined by computing the Low-Point value, as given
in Algorithm and illustrated on a
graph in .From the low-point value and lowpoint parent, there are two very
useful things which motivate our computation.First, if there is a child c of x such that L(c) >= D(x), then
there are no paths in the network graph that go from c or its
descendants to an ancestor of x - and therefore x is a cut-vertex.
This is useful because it allows identification of the cut-vertices
and thus the blocks. As seen in , even if L(x) < D(x), there
may be a block that contains both the root and a DFS-child of a node
while other DFS-children might be in different blocks. In this
example, C's child D is in the same block as R while F is not.Second, by repeatedly following the path given by lowpoint_parent,
there is a path from x back to an ancestor of x that does not use the
link [x, x.dfs_parent] in either direction. The full path need not be
taken, but this gives a way of finding an initial cycle and then
ears. A key idea for the MRT algorithm is that any non-2-connected graph
is made up by blocks (e.g. 2-connected clusters, cut-links, and/or
isolated nodes). To compute GADAGs and thus MRTs, computation is done
in each block to compute ADAGs or Redundant Trees and
then those ADAGs or Redundant Trees are combined into a GADAG or
MRT.Consider the example depicted in (a). In this figure, a special
graph is presented, showing us all the ways 2-connected clusters can
be connected. It has four blocks: block 1 contains R, A, B, C, D, E,
block 2 contains C, F, G, H, I, J, block 3 contains K, L, M, N, O, P,
and block 4 is a cut-edge containing H and K. As can be observed, the
first two blocks have one common node (node C) and blocks 2 and 3 do
not have any common node, but they are connected through a cut-edge
that is block 4. No two blocks can have more than one common node,
since two blocks with at least 2 common nodes would qualify as a
single 2-connected cluster.Moreover, observe that if we want to get from one block to another,
we must use a cut-vertex (the cut-vertices in this graph are C, H, K),
regardless of the path selected, so we can say that all the paths from
block 3 along the MRTs rooted at R will cross K first. This
observation means that if we want to find a pair of MRTs rooted at R,
then we need to build up a pair of RTs in block 3 with K as a
root. Similarly, we need to find another one in block 2 with C as a
root, and finally, we need the last one in block 1 with R as a
root. When all the trees are selected, we can simply combine them;
when a block is a cut-edge (as in block 4), that cut-edge is added in
the same direction to both of the trees. The resulting trees are
depicted in (b) and
(c).Similarly, to create a GADAG it is sufficient to compute ADAGs in
each block and connect them.It is necessary, therefore, to identify the cut-vertices, the
blocks and identify the appropriate local-root to use for each
block.Each node in a network graph has a local-root, which is the
cut-vertex (or root) in the same block that is closest to the root.
The local-root is used to determine whether two nodes share a common
block. There are two different ways of computing the local-root for each
node. The stand-alone method is given in and better illustrates the concept.
It is used in the second and third options for computing a GADAG using SPFs
and the hybrid versions respectively. The other method for local-root
computation is used in the first option for computing a GADAG using
Low-Point inheritance and the essence of it is given in .
Once the local-roots are known, two nodes X and Y are in a common
block if and only if one of the following three conditions apply.Y's local-root is X's local-root : They are in the same block and
neither is the cut-vertex closest to the root.Y's local-root is X: X is the cut-vertex closest to the root for
Y's blockY is X's local-root: Y is the cut-vertex closest to the root for
X's blockOnce we have computed the local-root for each node in the network
graph, we can assign for each node, a block id that represents the
block in which the node is present. This computation is
shown in . The block id is useful in the ear
computations involved in the SPF and hybrid based GADAG's as will be seen later.
This algorithm computes one GADAG that is then used by a router to
determine its blue MRT and red MRT next-hops to all destinations.
Finally, based upon that information, alternates are selected for each
next-hop to each destination. The different parts of this algorithm
are described below. These work on a network graph after, for
instance, its interfaces are ordered as per .Select the root to use for the GADAG. [See .]Initialize all interfaces to UNDIRECTED. [See .]Compute the DFS value,e.g. D(x), and lowpoint value, L(x). [See
.]Construct the GADAG. [See for
Option 1 using Lowpoint Inheritance,
for Option 2 using SPFs and
for Option 3 using a hybrid method.]Assign directions to all interfaces that are still UNDIRECTED. [See
.]From the computing router x, compute the next-hops for the blue MRT
and red MRT. [See .]Identify alternates for each next-hop to each destination
by determining which one of the blue MRT and the red MRT the computing
router x should select. [See .]To ensure consistency in computation, it is necessary that all
routers order interfaces identically. This is necessary for the DFS,
where the selection order of the interfaces to explore results in
different trees, and for computing the GADAG, where the selection
order of the interfaces to use to form ears can result in different
GADAGs. The recommended ordering between two interfaces from the same
router x is given in .The precise mechanism by which routers advertise a priority for the
GADAG root is not described in this document. Nor is the algorithm
for selecting routers based upon priority described in this
document.A network may be partitioned or there may be islands of routers
that support MRT fast-reroute. Therefore, the root selected for use
in a GADAG must be consistent only across each connected island of MRT
fast-reroute support. Before beginning computation, the network graph
is reduced to contain only the set of routers that support a
compatible MRT fast-reroute.The selection of a GADAG root is done among only those routers in
the same MRT fast-reroute island as the computing router x.
Additionally, only routers that are not marked as unusable or
overloaded (e.g. ISIS overload or ) are
eligible for selection as root.Before running the algorithm, there is the standard type of
initialization to be done, such as clearing any computed DFS-values,
lowpoint-values, DFS-parents, lowpoint-parents, any MRT-computed
next-hops, and flags associated with algorithm.It is assumed that a regular SPF computation has been run so that
the primary next-hops from the computing router to each destination
are known. This is required for determining alternates at the last
step.Initially, all interfaces must be initialized to UNDIRECTED.
Whether they are OUTGOING, INCOMING or both is determined when the
GADAG is constructed and augmented.It is possible that some links and nodes will be marked as
unusable, whether because of configuration, overload, or due to a
transient cause such as . In the algorithm
description, it is assumed that such links and nodes will not be
explored or used and no more disussion is given of this
restriction.The basic idea of this is to find ears from a node x that is
already in the GADAG (known as IN_GADAG). There are two methods to
find ears; both are required. The first is by going to a not IN_GADAG
DFS-child and then following the chain of low-point parents until an
IN_GADAG node is found. The second is by going to a not IN_GADAG
neighbor and then following the chain of DFS parents until an IN_GADAG
node is found. As an ear is found, the associated interfaces are
marked based on the direction taken. The nodes in the ear are marked
as IN_GADAG. In the algorithm, first the ears via DFS-children are
found and then the ears via DFS-neighbors are found.By adding both types of ears when an IN_GADAG node is processed,
all ears that connect to that node are found. The order in which the
IN_GADAG nodes is processed is, of course, key to the algorithm. The
order is a stack of ears so the most recent ear is found at the top of
the stack. Of course, the stack stores nodes and not ears, so an
ordered list of nodes, from the first node in the ear to the last node
in the ear, is created as the ear is explored and then that list is
pushed onto the stack.Each ear represents a partial order (see ) and processing the nodes in order along each
ear ensures that all ears connecting to a node are found before a node
higher in the partial order has its ears explored. This means that
the direction of the links in the ear is always from the node x being
processed towards the other end of the ear. Additionally, by using a
stack of ears, this means that any unprocessed nodes in previous ears
can only be ordered higher than nodes in the ears below it on the
stack.In this algorithm that depends upon Low-Point inheritance, it is
necessary that every node have a low-point parent that is not itself.
If a node is a cut-vertex, that may not yet be the case. Therefore,
any nodes without a low-point parent will have their low-point parent
set to their DFS parent and their low-point value set to the DFS-value
of their parent. This assignment also properly allows an ear between
two cut-vertices.Finally, the algorithm simultaneously computes each node's
local-root, as described in .
This is further elaborated as follows. The local-root can be
inherited from the node at the end of the ear unless the end of the ear
is x itself, in which case the local-root for all the nodes in the ear
would be x. This is because whenever the first cycle is found in a block,
or an ear involving a bridge is computed, the cut-vertex closest to the
root would be x itself. In all other scenarios, the properties of
lowpoint/dfs parents ensure that the end of the ear will be in the same block,
and thus inheriting its local-root would be the correct local-root for all
newly added nodes. The pseudo-code for the GADAG algorithm (assuming that the
adjustment of lowpoint for cut-vertices has been made) is shown in
.The basic idea in this option is to use slightly-modified SPF
computations to find ears. In every block, an SPF
computation is first done to find a cycle from the local root and then
SPF computations in that block find ears until there are no more interfaces
to be explored. The used result from the SPF computation is the path of
interfaces indicated by following the previous hops from the mininized
IN_GADAG node back to the SPF root.To do this, first all cut-vertices must be identified and local-roots
assigned as specified in .The slight modifications to the SPF are as follows. The root of the
block is referred to as the block-root; it is either the GADAG root or
a cut-vertex.The SPF is rooted at a neighbor x of an IN_GADAG node y. All links
between y and x are marked as TEMP_UNUSABLE. They should not be used
during the SPF computation.If y is not the block-root, then it is marked TEMP_UNUSABLE. It
should not be used during the SPF computation. This prevents ears
from starting and ending at the same node and avoids cycles; the
exception is because cycles to/from the block-root are acceptable
and expected.Do not explore links to nodes whose local-root is not the block-root.
This keeps the SPF confined to the particular block.Terminate when the first IN_GADAG node z is minimized.Respect the existing directions (e.g. INCOMING, OUTGOING,
UNDIRECTED) already specified for each interface.Assume that an ear is found by going from y to x and then running
an SPF that terminates by minimizing z
(e.g. y<->x...q<->z). Now it is necessary to determine
the direction of the ear; if y << z, then the path should be
y->x...q->z but if y >> z, then the path should be
y<-x...q<-z. In , the same
problem was handled by finding all ears that started at a node before
looking at ears starting at nodes higher in the partial order. In
this algorithm, using that approach could mean that new ears aren't
added in order of their total cost since all ears connected to a node
would need to be found before additional nodes could be found.The alternative is to track the order relationship of each node
with respect to every other node. This can be accomplished by
maintaining two sets of nodes at each node. The first set,
Higher_Nodes, contains all nodes that are known to be ordered above
the node. The second set, Lower_Nodes, contains all nodes that are
known to be ordered below the node. This is the approach used in this
algorithm.A goal of the algorithm is to find the shortest cycles and ears.
An ear is started by going to a neighbor x of an IN_GADAG node y. The
path from x to an IN_GADAG node is minimal, since it is computed via
SPF. Since a shortest path is made of shortest paths, to find the
shortest ears requires reaching from the set of IN_GADAG nodes to the
closest node that isn't IN_GADAG. Therefore, an ordered tree is
maintained of interfaces that could be explored from the IN_GADAG
nodes. The interfaces are ordered by their characteristics of metric,
local loopback address, remote loopback address, and ifindex, as in
the algorithm previously described in .The algorithm ignores interfaces picked from the ordered tree that belong
to the block root if the block in which the interface is present already has
an ear that has been computed. This is necessary since we allow at most one
incoming interface to a block root in each block. This requirement stems from
the way next-hops are computed as will be seen in
. After any ear gets computed, we
traverse the newly added nodes to the GADAG and insert interfaces whose far
end is not yet on the GADAG to the ordered tree for later processing.Finally, cut-edges are a special case because there is no point in
doing an SPF on a block of 2 nodes. The algorithm identifies
cut-edges simply as links where both ends of the link are
cut-vertices. Cut-edges can simply be added to the GADAG with both
OUTGOING and INCOMING specified on their interfaces.In this option, the idea is to combine the salient features of the above
two options. To this end, we process nodes as they get added to the GADAG just
like in the lowpoint inheritance by maintaining a stack of nodes. This
ensures that we do not need to maintain lower and higher sets at each node
to ascertain ear directions since the ears will always be directed from the
node being processed towards the end of the ear. To compute the ear however,
we resort to an SPF to have the possibility of better ears (path lentghs) thus
giving more flexibility than the restricted use of lowpoint/dfs parents.Regarding ears involving a block root, unlike the SPF method which ignored
interfaces of the block root after the first ear, in the hybrid method we
would have to process all interfaces of the block root before moving on to
other nodes in the block since the direction of an ear is pre-determined.
Thus, whenever the block already has an ear computed, and we are processing an
interface of the block root, we mark the block root as unusable before the SPF
run that computes the ear. This ensures that the SPF terminates at some node
other than the block-root. This in turn guarantees that the block-root has
only one incoming interface in each block, which is necessary for correctly
computing the next-hops on the GADAG. As in the SPF gadag, bridge ears are handled as a special case.The entire algorithm is shown below in The GADAG, whether constructed via Low-Point Inheritance or with
SPFs or the hybrid method, at this point could be used to find MRTs
but the topology does not include all links in the network graph.
That has two impacts.
First, there might be shorter paths that respect the GADAG partial
ordering and so the alternate paths would not be as short as possible.
Second, there may be additional paths between a router x and the root
that are not included in the GADAG. Including those provides
potentially more bandwidth to traffic flowing on the alternates and
may reduce congestion compared to just using the GADAG as currently
constructed.The goal is thus to assign direction to every remaining link marked
as UNDIRECTED to improve the paths and number of paths found when the
MRTs are computed.To do this, we need to establish a total order that respects the
partial order described by the GADAG. This can be done using Kahn's
topological sort which essentially
assigns a number to a node x only after all nodes before it (e.g. with
a link incoming to x) have had their numbers assigned. The only issue
with the topological sort is that it works on DAGs and not ADAGs or
GADAGs.To convert a GADAG to a DAG, it is necessary to remove all links
that point to a root of block from within that block. That provides
the necessary conversion to a DAG and then a topological sort can be
done. Finally, all UNDIRECTED links are assigned a direction based
upon the partial ordering. Any UNDIRECTED links that connect to a
root of a block from within that block are assigned a direction
INCOMING to that root. The exact details of this whole process are
captured in Proxy-nodes are used to represent multi-homed prefixes and routers
that do not support MRT Fast-Reroute. Until now, the network graph
has not included proxy-nodes because the computation for a GADAG
assumes that the nodes can be transited.To handle destinations that can only be reached via proxy-nodes,
each proxy-node should be added into the network graph after
Add_Directed_Links() has beeen run once. A proxy-node P is connected
to two routers, X and Y, which have been found to offer the best cost.
If X.topo_order < Y.topo_order, then the proxy-node P is added
along with a link X->P and a link P->Y. Once all the
proxy-nodes have been added in this fashion,
Run_Topological_Sort_GADAG() should be rerun so that the topological
order includes the proxy-nodes as well. This is needed for
determining which MRT can offer alternates, as is explained in .As was discussed in , once a ADAG
is found, it is straightforward to find the next-hops from any node X
to the ADAG root. However, in this algorithm, we want to reuse the
common GADAG and find not only one pair of redundant trees with it,
but a pair rooted at each node. This is ideal, since it is faster and
it results packet forwarding easier to trace and/or debug. The method
for doing that is based on two basic ideas. First, if two nodes X and
Y are ordered with respect to each other in the partial order, then
the same SPF and reverse-SPF can be used to find the increasing and
decreasing paths. Second, if two nodes X and Y aren't ordered with
respect to each other in the partial order, then intermediary nodes
can be used to create the paths by increasing/decreasing to the
intermediary and then decreasing/increasing to reach Y.As usual, the two basic ideas will be discussed assuming the
network is two-connected. The generalization to multiple blocks is
discussed in . The
full algorithm is given in .To find two node-disjoint paths from the computing router X to any
node Y, depends upon whether Y >> X or Y << X. As shown
in , if Y >> X, then there is an
increasing path that goes from X to Y without crossing R; this
contains nodes in the interval [X,Y]. There is also a decreasing path
that decreases towards R and then decreases from R to Y; this contains
nodes in the interval [X,R-small] or [R-great,Y]. The two paths
cannot have common nodes other than X and Y.Similar logic applies if Y << X, as shown in . In this case, the increasing path from X
increases to R and then increases from R to Y to use nodes in the
intervals [X,R-great] and [R-small, Y]. The decreasing path from X
reaches Y without crossing R and uses nodes in the interval [Y,X].When X and Y are not ordered, the first path should increase until
we get to a node G, where G >> Y. At G, we need to decrease to
Y. The other path should be just the opposite: we must decrease until
we get to a node H, where H << Y, and then increase. Since R is
smaller and greater than Y, such G and H must exist. It is also easy
to see that these two paths must be node disjoint: the first path
contains nodes in interval [X,G] and [Y,G], while the second path
contains nodes in interval [H,X] and [H,Y]. This is illustrated in
. It is necessary to decrease and
then increase for the Blue MRT and increase and then decrease for the
Red MRT; if one simply increased for one and decreased for the other,
then both paths would go through the root R.This gives disjoint paths as long as G and H are not the same node.
Since G >> Y and H << Y, if G and H could be the same
node, that would have to be the root R. This is not possible because
there is only one incoming interface to the root R which is created
when the initial cycle is found. Recall from that whenever an ear was found to have an
end that was the root R, the ear was directed from R so that the
associated interface on R is outgoing and not incoming. Therefore,
there must be exactly one node M which is the largest one before R, so
the Red MRT path will never reach R; it will turn at M and decrease
to Y.The basic ideas for computing RT next-hops in a 2-connected graph
were given in and . Given these two ideas, how can we
find the trees?If some node X only wants to find the next-hops (which is usually
the case for IP networks), it is enough to find which nodes are
greater and less than X, and which are not ordered; this can be done
by running an SPF and a reverse-SPF rooted at X and not exploring any
links from the ADAG root. ( Other traversal algorithms could safely
be used instead where one traversal takes the links in their given
directions and the other reverses the links' directions.)An SPF rooted at X and not exploring links from the root will find
the increasing next-hops to all Y >> X. Those increasing
next-hops are X's next-hops on the Blue MRT to reach Y. A reverse-SPF
rooted at X and not exploring links from the root will find the
decreasing next-hops to all Z << X. Those decreasing next-hops
are X's next-hops on the Red MRT to reach Z. Since the root R is both
greater than and less than X, after this SPF and reverse-SPF, X's
next-hops on the Blue MRT and on the Red MRT to reach R are known.
For every node Y >> X, X's next-hops on the Red MRT to reach Y
are set to those on the Red MRT to reach R. For every node Z <<
X, X's next-hops on the Blue MRT to reach Z are set to those on the
Blue MRT to reach R.For those nodes, which were not reached, we have the next-hops as
well. The increasing Blue MRT next-hop for a node, which is not
ordered, is the next-hop along the decreasing Red MRT towards R and
the decreasing Red MRT next-hop is the next-hop along the increasing
Blue MRT towards R. Naturally, since R is ordered with respect to all
the nodes, there will always be an increasing and a decreasing path
towards it. This algorithm does not provide the specific path taken
but only the appropriate next-hops to use. The identity of G and H is
not determined.The final case to considered is when the root R computes its own
next-hops. Since the root R is << all other nodes, running an
SPF rooted at R will reach all other nodes; the Blue MRT next-hops are
those found with this SPF. Similarly, since the root R is >>
all other nodes, running a reverse-SPF rooted at R will reach all
other nodes; the Red MRT next-hops are those found with this
reverse-SPF.As an example consider the situation depicted in . There node C runs an SPF and a
reverse-SPF The SPF reaches D, E and R and the reverse SPF reaches B,
A and R. So we immediately get that e.g. towards E the increasing
next-hop is D (it was reached though D), and the decreasing next-hop
is B (since R was reached though B). Since both D and B, A and R will
compute the next hops similarly, the packets will reach E.We have the next-hops towards F as well: since F is not ordered
with respect to C, the increasing next-hop is the decreasing one
towards R (which is B) and the decreasing next-hop is the increasing
one towards R (which is D). Since B is ordered with F, it will find a
real increasing next-hop, so packet forwarded to B will get to F on
path C-B-F. Similarly, D will have a real decreasing next-hop, and
packet will use path C-D-F.If a graph isn't 2-connected, then the basic approach given in
needs some extensions to
determine the appropriate MRT next-hops to use for destinations
outside the computing router X's blocks. In order to find a pair of
maximally redundant trees in that graph we need to find a pair of RTs
in each of the blocks (the root of these trees will be discussed
later), and combine them.When computing the MRT next-hops from a router X, there are three
basic differences:Only nodes in a common block with X should be explored in the SPF
and reverse-SPF.Instead of using the GADAG root, X's local-root should be used.
This has the following implications:
The links from X's local-root should not be explored. If a node is explored in the increasing SPF so Y
>> X, then X's Red MRT next-hops to reach Y uses X's Red MRT
next-hops to reach X's local-root and if Z <<, then X's Blue MRT
next-hops to reach Z uses X's Blue MRT next-hops to reach X's
local-root.If a node W in a common block with X was not reached in the SPF
or reverse-SPF, then W is unordered with respect to X. X's Blue MRT
next-hops to W are X's decreasing aka Red MRT next-hops to X's
local-root. X's Red MRT next-hops to W are X's increasing aka Blue
MRT next-hops to X's local-root.For nodes in different blocks, the next-hops must be inherited
via the relevant cut-vertex.These are all captured in the detailed algorithm given in .The complete algorithm to compute MRT Next-Hops for a particular
router X is given in . In
addition to computing the Blue MRT next-hops and Red MRT next-hops
used by X to reach each node Y, the algorithm also stores an
"order_proxy", which is the proper cut-vertex to reach Y if it is
outside the block, and which is used later in deciding whether the
Blue MRT or the Red MRT can provide an acceptable alternate for a
particular primary next-hop. At this point, a computing router S knows its Blue MRT next-hops and
Red MRT next-hops for each destination. The primary next-hops along
the SPT are also known. It remains to determine for each primary
next-hop to a destination D, which of the MRTs avoids the primary
next-hop node F. This computation depends upon data set in
Compute_MRT_NextHops such as each node y's y.blue_next_hops,
y.red_next_hops, y.order_proxy, y.higher, y.lower and topo_orders.
Recall that any router knows only which are the nodes greater and
lesser than itself, but it cannot decide the relation between any two
given nodes easily; that is why we need topological ordering.For each primary next-hop node F to each destination D, S can call
Select_Alternates(S, D, F, primary_intf) to determine whether to use
the Blue MRT next-hops as the alternate next-hop(s) for that primary next
hop or to use the Red MRT next-hops. The algorithm is given in and discussed afterwards.If either D>>S>>F or D<<S<<F holds true, the
situation is simple: in the first case we should choose the increasing Blue
next-hop, in the second case, the decreasing Red next-hop is the right
choice.However, when both D and F are greater than S the situation is not so
simple, there can be three possibilities: (i) F>>D (ii) F<<D or
(iii) F and D are not ordered. In the first case, we should choose the
path towards D along the Blue tree. In contrast, in case (ii) the
Red path towards the root and then to D would be the solution.
Finally, in case (iii) both paths would be acceptable. However,
observe that if e.g. F.topo_order>D.topo_order, either case (i) or
case (iii) holds true, which means that selecting the Blue next-hop
is safe. Similarly, if F.topo_order<D.topo_order, we should select
the Red next-hop. The situation is almost the same if both F and D
are less than S. Recall that we have added each link to the GADAG in some direction,
so it is impossible that S and F are not ordered. But it is
possible that S and D are not ordered, so we need to deal with this
case as well. If F<<S, we can use the Red next-hop, because that
path is first increasing until a node definitely greater than D is
reached, than decreasing; this path must avoid using F. Similarly, if
F>>S, we should use the Blue next-hop.Additionally, the cases where either F or D is ordered both higher
and lower must be considered; this can happen when one is a block-
root or inherits its order_proxy is. If D is both higher and lower
than S, then the MRT to use is the one that avoids F so if F is
higher, then the Red MRT should be used and if F is lower, then the
Blue MRT should be used; F and S must be ordered because they are
neighbors. If F is both higher and lower, then if D is lower, using
the Red MRT to decrease reaches D and if D is higher, using the Blue
MRT to increase reaches D; if D is unordered compared to S, then the
situation is a bit more complicated.
In the case where F<<S<<F and D and S are unordered, the
direction of the link in the GADAG between S and F should be
examined. If the link is directed S → F, then use the Blue MRT
(decrease to avoid that link and then increase). If the link is
directed S ← F, then use the Red MRT (increase to avoid that link
and then decrease). If the link is S ←→ F, then the link must be
a cut-link and there is no node-protecting alternate. If there are
multiple links between S and F, then they can protect against each
other; of course, in this situation, they are probably already ECMP.
Finally, there is the case where D is also F. In this case, only link
protection is possible. The MRT that doesn't use the indicated
primary next-hop is used. If both MRTs use the primary next-hop,
then the primary next-hop must be a cut-edge so either MRT could be
used but the set of MRT next-hops must be pruned to avoid that
primary next-hop. To indicate this case, Select_Alternates returns
AVOID_LINK_ON_BLUE.As an example, consider the ADAG depicted in
and first suppose that G is the
source, D is the destination and H is the failed next-hop.
Since D>>G, we need to compare H.topo_order and D.topo_order.
Since D.topo_order>H.topo_order, D must be not smaller than H, so
we should select the decreasing path towards the root. If, however, the
destination were instead J, we must find that H.topo_order>J.topo_order,
so we must choose the increasing Blue next-hop to J, which is I.
In the case, when instead the destination is C, we find that we need
to first decrease to avoid using H, so the Blue, first decreasing then
increasing, path is selected.This description of the algorithm assumes a particular approach
that is believed to be a reasonable compromise between complexity and
computation. There are two options given for constructing the GADAG
as both are reasonable and promising.Compute the common GADAG using Option 2
of SPF-based inheritance. This considers metrics when constructing
the GADAG, which is important for path length and operational control.
It has higher computational complexity than the Low-Point Inheritance
GADAG.Compute the common GADAG
using Option 1 of Low-Point Inheritance. This ignores metrics when
constructing the GADAG, but its computational complexity is O(links)
which is attractive. It is possible that augmenting the GADAG by
assigning directions to all links in the network graph and adding them
to the GADAG will make the difference between this and the SPF-based
GADAG minimal.In addition, it is possible to calculate Destination-Rooted GADAG,
where for each destination, a GADAG rooted at that destination is
computed. The GADAG can be computed using either Low-Point
Inheritance or SPF-based. Then a router would need to compute the
blue MRT and red MRT next-hops to that destination. Building GADAGs
per destination is computationally more expensive, but may give
somewhat shorter alternate paths. It may be useful for live-live
multicast along MRTs.When evaluating different algorithms and methods for IP Fast
Reroute , there are three critical points to consider.
Coverage: For every Point of Local Repair (PLR) and local failure,
is there an alternate to reach every destination? Those destinations
include not only routers in the IGP area, but also prefixes outside
the IGP area.Alternate Length: What is the length of the alternate path offered
compared to the optimal alternate route in the network? This is
computed as the total length of the alternate path divided by the
length of an optimal alternate path. The optimal alternate path is
computed by removing the failed node and running an SPF to find the
shortest path from the PLR to the destination.Alternate Bandwidth: What percentage of the traffic sent to the
failed point can be sent on the alternates? This is computed as the
sum of the bandwidths along the alternate paths divided by the
bandwidth of the primary paths that go through the failure point.Simulation and modeling to evalute the MRT algorithms is underway.
The algorithms being compared are:
SPF-based GADAGLow-Point Inheritance GADAGDestination-Rooted SPF-based GADAGDestination-Rooted Low-Point Inheritance GADAGNot-Via to Next-Next HopLoop-Free AlternatesRemote LFAsThe algorithm assumes that
broadcast interfaces are already represented as pseudo-nodes in the
network graph. The exact rules for extending the set of next-hops and
ensuring that the neighboring node is avoided need to be fully specified.The algorithmic extensions to
handle local SRLGs, where each member of the SRLG shares a common
router end, need to be fully specified.Creating MRTs that consider
general SRLGs is still a challenging open research problem.This doument includes no request to IANA.This architecture is not currently believed to introduce new security concerns.
&I-D.ietf-rtgwg-mrt-frr-architecture;
&RFC3137;
&RFC5286;
&RFC5714;
&I-D.ietf-rtgwg-ipfrr-notvia-addresses;
&I-D.shand-remote-lfa;
Topological sorting of large networksNovel Algorithms for IP Fast Reroute
&I-D.ietf-rtgwg-lfa-applicability;
IP Fast ReRoute: Lightweight Not-Via without Additional AddressesIP Fast ReRoute: Loop Free Alternates RevisitedOn Finding Maximally Redundant Trees in Strictly Linear Time