Content uploaded by Paulo Monteiro
Author content
All content in this area was uploaded by Paulo Monteiro on Feb 02, 2015
Content may be subject to copyright.
Efficient Multi-Path Routing for Optical Burst-Switched Networks
João Pedro1,2*, Jorge Castro1, Paulo Monteiro1,3, and João Pires2
1 D1 Research, Siemens Networks S.A., Rua Irmãos Siemens 1, 2720-093 Amadora, Portugal
2 Instituto de Telecomunicações, Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal
3 Instituto de Telecomunicações, Universidade de Aveiro, Campus Universitário de Santiago, 3810-193 Aveiro, Portugal
* phone: +351-214167179, fax: +351-214242082, e-mail: joao.pedro@siemens.com
Abstract1— Optical Burst Switching (OBS) is a promising
switching paradigm to efficiently support Internet Protocol (IP)
traffic over optical networks, under the current and foreseeable
limitations of optical technology. OBS network performance is
mainly hampered at the network nodes by resource contention
between bursts directed to the same output link. This paper
presents a multi-path routing strategy for minimizing contention
in OBS networks. Simulation results show that the proposed
routing strategy reduces burst blocking probability as compared
with single-path routing strategies.
I. INTRODUCTION
The Optical Burst Switching paradigm [1] has attracted
considerable interest as an optical networking architecture for
efficiently supporting IP packet traffic, while exploiting the
huge transmission capacity available with optical fibers and
Wavelength Division Multiplexing (WDM) technology [2].
OBS bandwidth granularity and technological complexity are
in between that of the Optical Circuit Switching (OCS) and
Optical Packet Switching (OPS) paradigms. The former
paradigm can be easily implemented with existing optical
technology, but it is inefficient in supporting IP traffic due to
its coarse wavelength granularity, whereas the later paradigm
provides statistical multiplexing at the packet level, but
requires optical buffering and optical processing capabilities,
which are still too immature to be deployed.
In OBS networks, multiple IP packets directed to the same
egress node are assembled at the ingress node to create a
burst. Since a wavelength is allocated to a burst only during
the time required for its transmission, this paradigm provides
sub-wavelength granularity. OBS technological requirements
are less stringent than those of OPS because, similarly to
OCS, bandwidth for data burst transmission is reserved in
advance and using out-of-band signaling, avoiding the need
for optical buffering and optical processing capabilities.
OBS networks use one-way resource reservation for setting
up the necessary resources for each burst transmission [1]. As
a result, the ingress node of the burst starts transmitting soon
after the burst has been assembled, instead of waiting for an
acknowledgment of successful resource reservation along the
entire burst path. Thus, bursts may contend for the same
resources at a transit node of the OBS network. Unresolved
contention leads to burst loss, which is the main cause of
performance degradation in OBS networks.
This work was supported by Siemens Networks S.A. and Fundação para a
Ciência e a Tecnologia (FCT) under research grant SFRH/BDE/15584/2006.
Resource contention at the network nodes can be resolved
using strategies acting in one or several of three domains:
wavelength, time, and space. In particular, the contending
bursts can be converted to other wavelengths using all-optical
wavelength converters, delayed using optical delay lines, or
deflected to other output links of the node. These strategies
are reactive by nature, attempting to resolve contention
instead of avoiding it. However, regardless of the contention
resolution strategies being used, the probability of resource
contention should be minimized a priori, whenever possible.
A primary source of contention is the congestion at the
bottleneck links of the OBS network, wherein a limited
number of wavelengths is shared by the largest amount of
burst traffic. Consequently, the first approach to minimize
resource contention consists of reducing the burst traffic load
offered to the network bottleneck links. In principle, this can
be achieved by optimizing the paths used to route the traffic.
In OCS networks, given the network-wide state information
availability, each traffic unit (circuit) can be routed by a least
congested path [3]. However, one-way resource reservation
and short burst duration, characteristic of OBS, prevent the
network nodes from maintaining updated network-wide state
information and, hence, using an optimal path per burst. Thus,
bursts are routed through the same primary paths, which are
defined as the paths used by default in the absence of burst
deflections at transit nodes. Eventually, the primary paths can
be modified as a response to changes in network and traffic
conditions, but taking place over relatively large time scales.
In this work, we propose using multi-path routing in OBS
networks to reduce congestion at the bottleneck links. In our
approach, a set of candidate paths is available to route bursts
between each node pair. The amount of burst traffic sent
through each path is determined using a Linear Programming
(LP) model. Two LP models are described. The first model
only minimizes the burst traffic load on the most congested
link, whereas the second model fine tunes the solution of the
first model such that the average offered traffic load per link
is also reduced. Moreover, a simple scheme for ordered burst
delivery is presented. The simulation results show that our
multi-path routing strategy reduces burst blocking probability
when compared with single-path routing strategies.
The remainder of the paper is organized as follows. Section
II overviews the routing strategies already proposed for OBS
networks. Section III describes a novel multi-path routing
strategy for minimizing resource contention at the core nodes
of OBS networks. The efficiency of this strategy in improving
network performance is evaluated through network simulation
in section IV. Finally, some concluding remarks are made in
section V.
II. ROUTING STRATEGIES FOR OBS NETWORKS
Routing path optimization in OBS networks has received
considerable attention in recent years [4]–[8]. Almost all of
these studies/proposals have focused on Single-Path Routing
(SPR) strategies, which define a unique primary path to route
bursts between each pair of nodes. The routing paths selection
is made either by monitoring a set of possible paths and
choosing a primary path according to the rate of successful
burst transmissions in the recent past [4], [5], or based on
long-term network and traffic information, such as network
topology and average offered traffic load values [6]–[8].
The routing strategies proposed in [4], [5] consider multiple
paths per node pair and select the path with the highest rate of
successful burst transmissions as the primary path. Computing
this rate requires the ingress nodes to receive feedback
messages either from the egress nodes, in case of successful
burst transmissions, or from the transit nodes, in case of burst
losses. These strategies have the advantage of being able to
adapt the primary path selection to changes in the traffic
pattern in relatively short times. The strategy in [4] monitors
the burst loss rate of each candidate path by sending a small
amount of data bursts through it. This may result in increased
burst loss as some of the traffic is being routed through paths
with large burst loss rate. Alternatively, the strategy in [5]
keeps track of the burst loss rate of each candidate path by
probing them each time a burst is transmitted, that is, search
packets are sent on the control channel to check whether the
burst, which is being sent through the primary path, would be
successfully routed on the other candidate paths. Thus, this
strategy significantly increases the bandwidth requirements of
the control channels.
The routing strategies proposed in [6]–[8] use as inputs the
network topology and average offered traffic load values to
determine a single primary path for each node pair. The work
in [6] uses a modified Dijkstra algorithm, whereas in [7] a
tabu-search based algorithm is used to optimize the primary
path selection. Finally, [8] proposes a more elaborated, and
rather complex, LP model to find the primary paths. In all
three strategies, solving the resulting optimization problem
requires the problem inputs to be available at some point of
the OBS network. The routing paths can be recomputed in
response to significant changes in the network topology or the
traffic pattern, as long as these modifications take place over
relatively long time scales.
III. MULTI-PATH ROUTING FOR OBS NETWORKS
Multi-Path Routing (MPR), wherein multiple primary paths
are used to route bursts between each node pair, has been
neglected in OBS networks due to studies that suggest that it
leads to increased burst loss at the transit nodes [4], as well as
out-of-order IP packet delivery at the egress nodes. In the
following, we propose an efficient multi-path routing strategy
for OBS networks and discuss a scheme for ordered IP packet
delivery at the egress nodes when using multi-path routing.
The proposed multi-path routing strategy for OBS networks
comprises two steps, namely: candidate path computation and
path load distribution. In the first step, a set of candidate
paths is computed for each node pair, whereas in the second
step the fraction of burst traffic routed through each candidate
path is determined. During network operation, the ingress
nodes will route the bursts generated locally through the
candidate paths according to this optimal load distribution. As
in [6]–[8], the long-term network and traffic information that
is assumed to be available for the routing strategy consists of
the network topology represented by a graph G(N, L), where
N is the set of network nodes and L is the set of network links,
and the average offered traffic load γsd for all Nds ∈
∈∈
∈,.
A. Candidate Path Computation
The motivation for decoupling candidate path computation
from path load distribution is mainly due to the following
advantages: limit the number of primary paths each ingress
node will have to handle, exclude in advance candidate paths
that are not feasible (for example, as a result of physical
impairments), and reduce the complexity of the optimization
problem. The main shortcoming of this approach is that the
routing problem solution may not be optimum, in the cases
where it will require paths that were eliminated from the set
of candidates in this step.
For simplicity, but without loss of generality, this work uses
a k-shortest path algorithm [9] to compute the set of candidate
paths Πsd for each Nds ∈
∈∈
∈, with γsd > 0. This algorithm has
been implemented in such way that it computes the k paths
with least number of hops, or all the paths with a number of
hops equal or smaller than h, or all paths under both limits.
B. Path Load Distribution
The path load distribution consists of determining the
fraction of burst traffic routed through each candidate path
that minimizes the burst traffic load offered to the bottleneck
link, which consequently reduces resource contention in the
OBS network. The problem can be formulated as a LP model
using the following notation. Let usd,p, l equal 1 if link Ll ∈
∈∈
∈ is
used by candidate path sd
pΠ∈
∈∈
∈, and 0 otherwise, and let xsd,p
denote the average traffic load offered to candidate path
sd
pΠ∈
∈∈
∈. Note that usd,p,l is defined by the candidate path
computation step. Moreover, let y denote the average offered
traffic load on the most congested (bottleneck) link. The
following LP model minimizes the burst traffic load offered
to the bottleneck link.
Minimize
y
(1)
subject to sd
p
psd
sd
xγ
,=
==
=
∑
∑∑
∑
∈
∈∈
∈Π
, Nds ∈
∈∈
∈
∀
∀∀
∀, (2)
0
,
,,, ≥
≥≥
≥⋅
⋅⋅
⋅−
−−
−
∑
∑∑
∑
∑
∑∑
∑
∈
∈∈
∈ ∈
∈∈
∈Nds p
lpsdpsd
sd
uxy
Π
, Ll∈
∈∈
∈
∀
∀∀
∀ (3)
The first set of constraints (2) is used to guarantee that the
entire traffic load offered between a pair of nodes is offered
to the available candidate paths, whereas the second set of
constraints (3) is used to obtain the average offered traffic
load on the bottleneck link. Using k candidate paths per node
pair and assuming that γsd > 0 for all Nds ∈
∈∈
∈, with s
≠
≠≠
≠
d, the
above LP model has |N|
×
××
×
|N-1|
×
××
×
k
+
++
+
1 variables, where |N| is
the number of network nodes, and |N|
×
××
×
|N-1|
+
++
+
|L| constraints,
where |L| is the number of network links. However, since all
problem variables are real numbers, the LP model is expected
to be promptly solved even for large-sized networks.
A potential limitation of this LP model lies on the fact that,
as the number of candidate paths per node pair is increased
and the minimum possible average offered traffic load on the
bottleneck link has already been achieved, it can produce
solutions which unnecessarily distribute burst traffic load on
longer paths, thus increasing resource utilization. In order to
avoid this, a second LP model which minimizes the average
offered traffic load per link, using the bottleneck link load
determined in the first LP model as a constraint is proposed.
Minimize ∑
∑∑
∑ ∑
∑∑
∑ ∑
∑∑
∑
∈
∈∈
∈ ∈
∈∈
∈ ∈
∈∈
∈
⋅
⋅⋅
⋅
Ll Nds p
lpsdpsd
sd
ux
L,
,,,
1
Π
(4)
subject to sd
p
psd
sd
xγ
,=
==
=
∑
∑∑
∑
∈
∈∈
∈Π
, Nds ∈
∈∈
∈
∀
∀∀
∀, (5)
yux
Nds p
lpsdpsd
sd
≤
≤≤
≤⋅
⋅⋅
⋅
∑
∑∑
∑
∑
∑∑
∑
∈
∈∈
∈ ∈
∈∈
∈,
,,,
Π
, Ll∈
∈∈
∈
∀
∀∀
∀ (6)
This LP model has the same number of constraints than the
former model and minus one variable. After solving the LP
model(s), only the candidate paths with xsd,p > 0 will be used
by ingress node s. Upon assembling a burst destined for node
d, the ingress node s will try to route the burst through path
sd
pΠ∈
∈∈
∈ with probability xsd,p /γsd.
As a final remark, note that this routing strategy, as most
strategies proposed, is based on the assumption that during
network operation burst loss is small (10-3 or less [10]) and,
hence, the load effectively submitted to a link, accounting for
burst losses at transit nodes, does not significantly differ from
the load observed if the links had unlimited capacity.
C. Ordered Burst Delivery
The use of multiple paths between two nodes may result in
occasional out-of-order burst delivery at the egress node, as a
burst which started to be transmitted earlier, but on a longer
path, may arrive at the egress node after a burst transmitted
later on a shorter path. This results in out-of-order IP packet
delivery, which may be intolerable in some IP-based services.
The following scheme is used here to avoid this effect. Let
tp be the burst transmission delay when using path p, which is
given by the sum of the offset time needed to transmit a burst
on p and the path propagation delay, and let tmax be the largest
transmission delay among paths of Πsd with γsd > 0. Ordered
burst delivery is guaranteed if ingress node s introduces an
additional delay of tmax–tp to bursts going through path p.
VI. RESULTS AND DISCUSSION
The performance of multi-path routing in OBS networks is
evaluated here using network simulation [10]. The metric
used is the average burst blocking probability, measuring the
average fraction of burst traffic lost by the network. The NSF
and COST 239 reference networks of Fig. 1 are used in this
study. The former assumes a uniform traffic pattern, whereas
the later has a non-uniform traffic pattern [7]. In both cases,
the OBS network employs Just Enough Time (JET) resource
reservation [1] and full-range wavelength conversion. In
addition, it has 64 wavelengths per link, 10 Gb/s wavelength
capacity, a switch fabric configuration time of 10 µs, an
average burst size of 100 kB, and exponentially distributed
burst size and burst interarrival time. The average offered
traffic load normalized to the network capacity is given by
BWL
Nds sd
×
××
××
××
×
=
==
=
∑
∑∑
∑
∈
∈∈
∈,γ
Γ, (7)
where W is the number of wavelengths per link and B is the
wavelength capacity. The LP models are solved using [11].
(a) NSF (b) COST 239
Fig. 1. Reference network topologies.
Fig. 2 plots the average burst blocking probability as a
function of the number of candidate paths per node pair (k)
for the COST 239 network with Γ=0.30 and 0.35 and using
both the LP model that only minimizes the offered traffic load
on the bottleneck link (I) and the LP model that minimizes the
average offered traffic load per link constrained to minimum
offered traffic load on the bottleneck link (II).
1.0E-4
1.0E-3
1.0E-2
1.0E-1
1.0E+0
1 2 3 4 5 6 7 8 9
Number of candidat e paths pe r node pair
Aver age burst blocking probability
LP M odel I
LP M odel II
10
-1
10
-3
10
-4
10
-2
10
0
Γ=0.35
Γ=0.30
Fig. 2. Average burst blocking probability using the LP models for
path load distribution in the COST 239 network.
The curves suggest that there is no significant improvement
in performance for more than 4 or 5 candidate paths per node
pair. Moreover, as expected, it is also noticeable that LP
model I produces worst solutions as the number of candidate
paths becomes very large because, unlike LP model II, it is
not able to avoid the unnecessary use of longer paths.
In the following, the performance of the MPR strategy using
the LP model II and k=5 is compared to that of two SPR
strategies, namely the Minimum Hop Count (MHC) and the
Heuristic Minimum Network Congestion (HMNC) [7]. Fig. 3
and Fig. 4 plot the average burst blocking probability as a
function of the normalized offered traffic load for the COST
239 and NSF networks, respectively.
1.0E-4
1.0E-3
1.0E-2
1.0E-1
1.0E+0
0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Average offe red tra ffic load
Ave rage burst blocking probability
SPR - MHC
SPR - HMNC
MPR
10
-1
10
-3
10
-4
10
-2
10
0
Fig. 3. Average burst blocking probability using single-path and
multi-path routing strategies in the COST 239 network.
1.0E-4
1.0E-3
1.0E-2
1.0E-1
1.0E+0
0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
Average offe red tra ffic load
Ave rage burst blocking probability
SPR - MHC
SPR - HMNC
MPR
10
-1
10
-3
10
-4
10
-2
10
0
Fig. 4. Average burst blocking probability using single-path and
multi-path routing strategies in the NSF network.
The plots show that both SPR-HMNC and MPR outperform
SPR-MHC, confirming that the minimization of the traffic
load offered to the bottleneck link improves OBS network
performance. In addition, since MPR always performs better
than SPR-HMNC, this suggests that multi-path routing can be
more efficient than single-path routing in OBS networks, as it
provides an additional degree of freedom for load balancing.
V. CONCLUSIONS
This paper has proposed a novel multi-path routing strategy
for minimizing contention in OBS networks by reducing the
congestion at the bottleneck links. The strategy decouples the
candidate path computation and the path load distribution to
increase flexibility and reduce computation time. Simulation
results show that multi-path routing reduces burst blocking
probability as compared with single-path routing.
Future work includes assessing the potential performance
improvement from combining the contention minimization
ability of our multi-path routing strategy with contention
resolution through deflection routing. In theory, this will fully
exploit the space domain for improving network performance.
REFERENCES
[1] C. Qiao, and M. Yoo, “Optical Burst Switching (OBS) –
A new paradigm for an optical Internet”, Journal of High
Speed Networks, vol. 8, no. 1, pp. 69-84, January 1999.
[2] R. Ramaswami, and K. Sivarajan, Optical Networks: A
practical perspective, San Francisco: Morgan Kaufmann,
2002.
[3] H. Zang, J. Jue, and B. Mukherjee, “A review of routing
and wavelength assignment approaches for wavelength-
routed optical WDM networks”, Optical Networks
Magazine, vol. 1, no. 1, pp. 47-60, January 2000.
[4] S. Ganguly, S. Bhatnagar, R. Izmailov, and C. Qiao,
“Multi-path optical burst forwarding”, in Proc. of IEEE
Workshop on High Performance Switching and Routing
2004, Phoenix, USA, April 2004, pp. 180-185.
[5] D. Ishii, N. Yamanaka, and I. Sasase, “A self-learning
route selection scheme using multi-path searching
packets in an OBS network”, Journal of Optical
Networking, vol. 4, no. 7, pp. 432-445, July 2005.
[6] Y. Du, T. Pu, H. Zhang, and Y. Guo, “Adaptive load
balancing routing algorithm for optical burst-switching
networks“, in Proc. of OFC 2005, Anaheim, USA, paper
OThF7.
[7] J. Castro, J. Pedro, and P. Monteiro, “Burst loss
reduction in OBS networks by minimizing network
congestion”, in Proc. of ConfTele 2005, Tomar, Portugal,
April 2005.
[8] J. Teng, and G. Rouskas, “Traffic engineering approach
to path selection in optical burst switching networks”,
Journal of Optical Networking, vol. 4, no. 11, pp. 759-
777, November 2005.
[9] E. Martins and M. Pascoal, ”An Algorithm for Ranking
Optimal Paths”, October 2000. Available:
http://www.mat.uc.pt/~eqvm/cientificos/investigacao/Arti
gos/rank_optimal.ps.gz
[10] J. Pedro, J. Castro, P. Monteiro, and J. Pires, “On the
modelling and performance evaluation of optical burst-
switched networks”, in Proc. of IEEE CAMAD 2006,
Trento, Italy, June 2006, pp. 30-37.
[11]
http://lpsolve.sourceforge.net/5.5/
