Simple Layout Algorithms To Maintain

Network Connectivity Under Faults

Galen H. Sasaki1
Department of Electrical, University of Hawaii
2540 Dole Street, Honolulu, HI 96822 USA

Ching-Fong Su and David Blight

Fujitsu Laboratories of America
595 Lawrence Expressway, Sunnyvale, CA 94086-3922 USA

Abstract: A problem of survivable layout of an IP network topology on a wavelength

division multiplexed network is considered. Simple layout algorithms are given that have
low time complexities. The algorithms are variations of a basic algorithm, that lays out IP
links as lightpaths sequentially according to some order and with respect to link costs.
Simulations show that the ordering and link costs can significantly affect the performance of
the algorithm.

1. Introduction

Wavelength division multiplexing (WDM) is the current technology of choice for broadband
backbone networks. WDM long-haul, transport systems are being deployed due to their low
cost per bit, good quality of service, and very wide bandwidths. WDM is essentially
frequency division multiplexing of optical signals onto optical fibers, and wavelengths are
equivalent to carrier frequencies (actually, the inverse of carrier frequencies). The
technology’s capability of providing very wide bandwidths using the existing optical fiber
resources make it a good match to meet the exponential growth of Internet traffic.

In this paper we consider the problem of supporting IP networks over WDM networks.
Figure 1 shows this two-layer network hierarchy, where the IP links are realized by optical
end-to-end connections, referred to as lightpaths, at the WDM layer. At the WDM layer are
fiber-pairs, each carrying multiple optical signals at different wavelengths. A fiber-pair
forms a full-duplex physical link, with the fibers carrying signals in opposite directions. The
fiber-pair carries full-duplex optical channels at each wavelength. In this paper, the fiber-
pairs will be referred to as fiber-links. The fiber-links are terminated by transport equipment
that multiplexes/demultiplexes the optical signals so that the fiber-link operates as a
collection of parallel, full-duplex optical channels at different wavelengths. Today,
lightpaths and WDM channels are 2.5 Gbps (OC-48) and 10 Gbps (OC-192), while 40 Gbps
(OC-768) is the emerging technology.

The WDM optical channels from the transport equipment may be cross-connected at optical
cross-connects (OXCs). The transport equipment has O-E-Os (optical-electrical-optical)
equipment, which implies that any pair of optical channels incident to a common OXC may
be cross-connected. (Note that without O-E-Os only channels with the same wavelength
may be cross-connected.)

1
The author was supported in part by Fujitsu Laboratories of America, and the National Science Foundation
under grant NCR-9612846.
Full-duplex lightpaths are formed at the WDM layer, by cross-connecting WDM channels at
OXCs. The lightpaths are terminated at interface ports of high speed IP routers, that are
connected to OXCs as shown in Figure 1. To the IP routers, lightpaths are their physical
links.

Since the IP network has high capacity, and thus supports many end-users, it is important that
it is survivable under faults. At a minimum, the network should remain connected under any
possible fault. As long as the IP network remains connected, it can reorganize its routing to
provide some minimum level of packet transport service, and its control plane will remain
operational.

IP Router

Lightpath C2 D2
realizing an
IP link Fiber-pair link
C1 D1

A2 A1 E1 E2
B1

Interface
ports OXC B2

Figure 1. IP network over WDM network.

The type of faults that we consider is the fiber-link failure (e.g., fiber cut). Thus, it is
necessary that the IP network topology is 2-connected, which means that the topology
remains connected under any single link failure. Though 2-connectivity is necessary it is not
sufficient. If lightpaths (realizing IP links) pass through a common fiber-link then a fiber-
link failure will result in multiple IP link failures. This may disconnect the IP network. For
example, consider Figure 1. The IP network topology is 2-connected. However, its lightpath
layout makes it vulnerable to single fiber-link cuts. In particular, if the fiber-link between
OXCs A1 and C1 fails then the two IP-links between IP router A2 and routers B2 and C2 fail,
leaving IP router A2 disconnected.

In this paper we consider the problem of laying out the IP network topology over the WDM
network so that the former is connected under single fiber-link faults. This was first
considered by Crochat and Le Boudec [1], and more recently in [2]. The time complexity of
their algorithm for solving the problem can be prohibitive for large networks. We propose
simpler algorithms that have lower time complexities.

In Section 2, we describe the network model and survivable layout problem. In Section 3,
heuristic layout algorithms with low time complexities are given. Simulation results
demonstrating the efficacy are presented in Section 4, and finally concluding remarks are
given in Section 5.
2. Network Model and Survivable Layout Problem

There are two network topologies: the IP and WDM network topologies. Both have the
same set of nodes V. Each node in V corresponds to an OXC and IP-router pair, where the
router is connected to the OXC with interface ports. The fiber-links between the OXCs form
the links EWDM of the WDM network. GWDM = (V, EWDM) denotes the WDM network
topology. The IP network topology is denoted by GIP = (V, EIP), where EIP is the collection
of IP links. An IP link is realized by a lightpath in the WDM network. A collection of
lightpaths that realize the IP links is referred to as a layout. We will assume that the WDM
network has a large number of wavelengths, so that the number of lightpaths that may
traverse a fiber-link is unrestricted.

The survivable layout problem is to find a layout for GIP such that any single fiber-link fault
in GWDM will leave the GIP topology connected. This problem is a simplified version of the
one considered by Crochat and Le Boudec [1, 2]. They consider the case where each layout
has a cost, and the problem is to find a layout with minimum cost. The cost is zero if and
only if the layout is survivable for any single fiber-link fault.

3. Heuristic Algorithms

Before describing our simple layout heuristics, we will briefly discuss the the algorithm of
Crochat and Le Boudec [1, 2] for the survivable layout problem. Their algorithm is a tabu
search, which is an iterative optimization algorithm that searches through possible solutions
to a problem. In this case, the possible solutions are lightpath layouts. Each solution has a
cost F, which in [1, 2] is a measure of how close it is to solving the survivable layout
problem. The cost F is the linear combination of three components: F1, F2, and F3. F1 is
the sum of fiber-link costs, where the cost of a link e is max (0, r (e) − C (e)) , C(e) is the
2

capacity of link e, and r(e) is the number of lightpaths on e. (In our case, C(e) is essentially
infinite since we assume a very large number of wavelengths.) This cost penalizes fiber-
links with load exceeding their link capacities. F2 is a measure of the number of IP links
broken for a single link fault. F3 is the sum of fiber-link costs, where the cost of a link e is
max (0, r (e) − t (e) ) and t(e) is some threshold. The threshold t(e) can be set to 1 to balance
2

lightpath loads across fiber-links.

During one of its iterations, tabu search determines a new layout Lnew from its current one
Lcurrent. It does this by first creating a collection of layouts that are “perturbations” of Lcurrent.
A perturbation is basically Lcurrent with exactly one of its lightpaths randomly rerouted. Then
Lnew is chosen among the perturbations with minimum. Note that each iteration requires |EIP|
perturbations, and each perturbation must have its cost computed. Crochat and Le Boudec
[1, 2] reported that each iteration takes O( |EWDM | |EIP | | V |2 ) time. The time complexity of
an iteration is dominated by the time it takes to evaluate the cost of a candidate layout and in
particular cost F2.

Thus, the algorithm does not scale well for large networks, especially if it requires many
iterations to find a good solution. To address the scalability problem, we investigate simpler
layout algorithms that have time complexities on the order of O(|EIP| |V |2). The complexity
is dominated by computing |EIP| lightpath routes, each requiring O(|V|2) shortest path
computation.
The algorithms we consider find lightpaths for subsets of IP links at a time. Once a lightpath
is computed for an IP link, the IP link is not considered again. The computation of the
lightpaths for a subset of size k has time complexity O(k |V|2), which is dominated by k
shortest path computations. Thus, the overall time complexity of the algorithms are O(|EIP|
|V |2).

Figure 2 shows the basic algorithm. The nodes are considered in some order. For each node
v, IP links incident to the node, which we denote by L(v), will have their lightpaths
computed. Only those IP links that do not yet have lightpaths will have their lightpaths
computed, i.e., an IP link will have its lightpath computed only once. Finding lightpaths for
the IP links in L(v) can be done by either routing them one-by-one using shortest path
routing, or together using minimum cost link-disjoint path routing [3]. One-by-one
computations are referred to as single route computations, while finding the routes together is
referred to as a multiple route computation. Multiple route computation is implemented by
the link-disjoint path routing of [3] and is described in Appendix A.

Survivable Layout Algorithm

1. Order the nodes. Denote the nodes by v(1), v(2),..., v(n), where n = |V| and
v(k) is the kth node.
2. For each k = 1, 2, ..., n,
2a. Let L(v(k)) be the collection of IP links incident to v(k).
2b. Layout lightpaths for the links in L(v(k)) that do not have
lightpaths. (Note that the layout can be done by single route
computations or by a multiple route computation.)

Figure 2. Basic algorithm to compute lightpath layout of an IP network topology.

Step 2b in Figure 2 is the step for single or multiple lightpath route computations. In the case
of single route computation, a lightpath is computed by first having costs assigned to fiber-
links of GWDM. Then a shortest path computation is applied to find the lightpath. In the case
of multiple route computation, first costs are assigned to fiber-links of GWDM. Then minimum
cost, link disjoint paths are computed (as described in Appendix A).

By appropriately choosing the costs, one can control the type of paths that will be found. In
the case of single route computations, the fiber-links have the following costs.

• Unit link (UNIT) cost: Each fiber-link has unit cost. This leads to shortest hop
lightpaths, which minimize the amount of bandwidth use. However, this may also
lead to an imbalance of lightpath loads on fiber-links. Such an imbalance should be
avoided because a fault along a heavily loaded fiber-link will result in faults on many
IP links.

• Quadratic load (QUAD) cost: Each fiber-link e has cost (le + 1)2, where le is the
number of lightpaths using e. This tends to distribute lightpaths evenly over fiber-
links, reducing multiple IP link failures per fiber-link failure. This is similar to cost
F3 of Crochat and Le Boudec [1, 2].
 • Existing Lightpaths-Single Route Computation (EL-S) cost: This cost depends on the
lightpath being computed. The fiber-link costs are as follows. Suppose that s and t
are the terminating nodes of the lightpath being computed. A fiber-link incident to s
(resp., t) will have high cost K (= 2 |EIP| |V|) if it carries a lightpath that terminates at s
(resp., t). The other links have unit cost. This tends to keep lightpaths that terminate
at a common node from using the same links.

As an example, consider Figure 1 with its existing collection of lightpaths. Suppose a

new lightpath is to be computed between routers at A2 and E2. Then the fiber-links
(A1,C1), (B1,E1), and (D1,E1) will have high cost K because they are incident to
OXC A1 or E1, and they carry lightpaths terminating at those OXCs.

In the case of multiple route computations, we use the following fiber-link cost.

• Existing Lightpaths-Multiple Route Computation (EL-M) cost: This cost depends on

node v(k) of Step 2 of the algorithm (see Figure 2). A fiber-link incident to v(k) that
has a lightpath terminating at v(k) has high cost K = 2 |EIP| |V| (same K, as in the EL-
S cost). The other fiber-links have unit cost. Like the EL-S cost, this tends to keep
lightpaths that terminate at a common node from using the same links.

As an example, consider Figure 1 with its existing collection of lightpaths. Suppose

new lightpaths were to be computed from the router A2. Since the existing lightpaths
that terminate at A2 use fiber-link (A1,C1), the link has cost K and the other links
have unit cost.

We should note that with multiple route computation, the survivable layout algorithm may
end up with some IP links without lightpaths. This can occur if there are many IP links, but
not many fiber-link disjoint paths out of nodes. To take care of the IP links without
lightpaths, the survivable lightpath layout algorithm is run again for these links but now for
single route computations and EL-S cost.

We consider two types of node orderings.

• Fixed ordering (FIXED): The ordering is arbitrary and fixed. In this paper, we order
according to node IDs.

• Smallest degree first (DEGR): This is an ordering according to node degrees, where
the degree of a node is the number of IP links incident to it. Presumably nodes with
smaller degrees are more likely to be disconnected from a fault, so they are
considered first.

4. Performance Results

To compare the options of the survivable lightpath layout algorithm in the previous section,
we ran simulations. For each option we ran simulations on three topologies of GWDM shown
in Figure 3. Topology G1 was created by hand, while topologies G2 and G3 were randomly
created. The random topologies were generated iteratively starting from initial topologies,
that have no fiber-links. At each iteration, a fiber-link is added between a random pair of
nodes, each pair being equally likely until the topology is 2-connected. However, we
disallow multiple links between nodes.

1
2
1 3 1 3 0

6
0 5 0 5

2 4 2 5 3
4
4
Topology G1 Topology G2
Topology G3

Figure 3. Three WDM network topologies.

For each WDM topology, 100 random IP topologies (GIP) were created. These topologies
were created iteratively in the same way topologies G2 and G3 were created. Each IP
topology has a layout computed for the WDM topology using the survivable layout
algorithm. Table 1 shows the results for single route computations when the node ordering is
FIXED or DEGR, and the fiber-link costs are UNIT, QUAD, or EL-S. There are results for
each WDM topology. Each table entry has the number of IP topologies that result in
survivable layouts. For example, for WDM topology G1, the number of IP topologies that
result in survivable layouts is 43 if the node order is FIXED and the link cost is QUAD.

WDM Network Topology

G1 G2 G3
Link Cost Link Cost Link Cost
Order UNIT QUAD EL-S UNIT QUAD EL-S UNIT QUAD EL-S
FIXED 33 43 70 33 50 48 34 75 84
DEGR 33 60 84 33 69 69 34 90 92

Table 1. Performance results for single route computation.

The table shows that node ordering and link costs can significantly improve performance. In
the case of node ordering, DEGR performs at least as well as FIXED, and in most cases the
performance difference is significant. In the case of link costs, EL-S performs best. Thus,
the combination of options DEGR and EL-S works best for single route computation.

Table 2 compares single and multiple route computation. The data for the single route
computation for the options DEGR and EL-S, which we have shown work best. The table
shows that multiple route computations performs about the same or better.
 WDM Network Topology
Routing G1 G2 G3
Single 84 69 92
Multiple 84 85 91

Table 2. Performance results for single versus multiple route computation.

5. Concluding Remarks

We propose simple algorithms to compute a survivable lightpath layout for an IP network

topology. It is an iterative approach that computes lightpaths in sequence. Link cost and
node ordering options were compared for single route computation. It was shown that the
choice of node ordering and link costs can result in significant improvement. In the case of
node ordering, it is better to first layout lightpaths of nodes with a small number of IP links
since they are more vulnerable to faults. In the case of link costs, UNIT is the worse since it
will not spread lightpaths over links. QUAD is better since the link costs tend to balance
lightpath load over links. EL-S is a little better than QUAD, perhaps because its link costs
tend to spread lightpath load and it is also specific to the lightpath being routed. We finally
compared single and multiple route computation, and found that multiple route computation
is about the same or better.

A. Link disjoint path algorithm

In this appendix we discuss how multiple route computations are done for IP links L(v),
which are the IP links incident to node v. The algorithm is basically a minimum cost, link
disjoint path algorithm of [3] on an appropriately modified network topology. The algorithm
will compute lightpaths for those IP links in L(v) that do not have one. Let n denote the
number of such IP links, and denote the links by L’(v) = { (v, u1), (v, u2), ...,(v, un) }.

The algorithm of [3] computes minimum cost, link disjoint paths in a graph between a pair of
nodes. If the number of disjoint paths is k then the algorithm’s time complexity is dominated
by the time complexity of k shortest path computations. Thus, the algorithm’s time
complexity is O(k |V|2). We apply the link disjoint path algorithm to a modified version of
network topology GWDM. In particular, a new (fictitious) node t is added. Also added are
new links (u1, t), (u2, t),..., (un, t), i.e., these are new links connecting the nodes of {u1, u2,...,
un} to node t. The new links have zero cost. Then the minimum cost, disjoint path algorithm
of [3] is applied to the modified topology GWDM, where the paths terminate at nodes v and t.
Note that the computed paths can be truncated at the nodes {u1, u2,..., un} to form link
disjoint lightpaths for a subset of the IP links in L’(v). Those IP links in L’(v) that do not get
a lightpath are left for a future lightpath computation.
References

1. O. Crochat, J.-Y. Le Boudec, “Design protection for WDM networks,” IEEE J. Selected
Areas on Communications, vol. 16, no. 7, pp. 1158-1165, Sept. 1998.
2. O. Crochat, J.-Y. Le Boudec, O. Gerstel, “Protection interoperability for WDM optical
networks,” unpublished manuscript. http://lrcwww.epfl.ch/crochat.
3. J. Surballe, “Disjoint paths in a network,” Networks, vol. 4, 1977, pp. 125-145.