Quality of Service Management in GMPLS-based Grid OBS
Universidade Federal do Pará
Rua Augusto Corrêa, 01
Belém, Brazil Belém, Brazil
Universidade Federal do Pará
Rua Augusto Corrêa, 01
Universidade Federal Fluminense
Rua Passo da Pátria, 156
Rio de Janeiro, Brazil
This paper proposes an architecture for the establishment of routes
with absolute QoS constraints for optical burst switched grid
networks. This model uses traffic engineering provided by
GMPLS to build LSPs that matches the required performance in
response to a request made by the user/application of the grid.
Results show that the proposal is capable of enforcing QoS by
reducing the loss experienced by burst classes and allowing a
better utilization of the computing resources.
Categories and Subject Descriptors
Operations – network management
Management, Design, Performance.
Optical Burst Switching, Quality of Service, Grid Networks,
The grid is a model that proposes the use of computational
resources from several machines situated in different locations to
solve problems that demand great computational power . In
addition to computers, a grid is composed of data repositories,
scientific instruments and visualization devices, amongst others,
which offer great opportunities for collaboration.
Computer networks are basic building blocks to for grid
construction, and efforts are being directed to developing
communication models more suited to the needs of grid
computing. One important aspect is the change in the way
computer networks are seen in the context of grid computing.
Until now, the network infrastructure that supports grid
computing has been considered an important component, although
not one fully integrated with other grid resources, and seen more
as an external component. Now, the tendency is to regard the
network as a grid resource, just like the processors, memory and
input/output devices, that provides new alternatives for service
Due to the large volume of data manipulated, a grid needs a robust
communication infrastructure that is appropriate to the
characteristics of this model. In particular, recent advances in
optical communications have been made in order to make all-
optical networks capable of supporting new network services,
such as grid applications, which may have strict performance
In this context, the all-optical switching paradigm, which attempts
to eliminate the existing limitations of electronic switching, and
its different alternatives (lambda switching, packet switching and
burst switching) appears as an alternative that turns these next-
generation optical networks into the best candidates for grid
computing. Specifically, optical burst switching (OBS) 
presents some advantages compared to the other all-optical
switching alternatives, like high link utilization and low
processing and synchronization overhead, and is easily integrated
with grid computing as jobs can be mapped to optical bursts .
A recurrent discussion is about the control plane used for the
functions of routing resource reservation and failure recovery,
among others. There are centralized approaches for a control
plane that are best suited for situations requiring a more precise
and flexible resource reservation, because in these centralized
planes we can have a complete view of all resources.
A distributed control plane is adequate when there is a need to
establish connections rapidly and with resource discovery
capabilities . The GMPLS (Generalized Multiprotocol Label
Switching) architecture  is a great option to build a distributed
control plane due to its suitability to optical networks, when the
labels are used to represent the wavelengths in the fiber, and
because of its support for traffic engineering. The current trend is
to consider a hybrid alternative, with the coexistence of
centralized and distributed functionalities, which would provide
broader support for the demands of grid computing.
The objective of this paper is to propose an architecture to
provision network routes that satisfy the performance constraints
of a grid job. This architecture is based on optical burst switching
and on the utilization of a control plane that combines GMPLS
signaling with a component that stores information about grid
resources (both computing and network resources). This
component, called the Grid OBS Connection Server, processes
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
SAC’09, March 8-12, 2009, Honolulu, Hawaii, U.S.A.
Copyright 2009 ACM 978-1-60558-166-8/09/03…$5.00.
requests about the availability of resources to handle a job and
calculates a possible route that satisfies the QoS constraints of the
request. With this information, the GMPLS signaling is used to
reserve all the resources along the path.
In addition to this introductory section, the paper is made up of
five other sections. Section 2 presents some concepts related to
optical burst switched networks in the context of grid computing,
section 3 describes the architecture proposed for LSP (Label
Switched Path) calculation in Grid OBS networks. The simulation
results are shown in section 4. Finally, section 5 mentions related
work and section 6 presents conclusions and some possibilities for
2. GRID OBS NETWORKS
Communication networks need to be adapted to grid computing
because they are fundamental to the viability of the services
offered by the grid. Many research initiatives are being led to
create new communication models for grid computing. At this
point, next-generation optical networks based on all-optical
switching and on the utilization of advanced control planes are
seen as major drivers to achieve this.
In the context of all-optical networks there are basically three
switching approaches: lambda switching, packet switching and
burst switching. Optical burst switching attempts to overcome the
limitations of the other alternatives like low resource utilization
and high implementation complexity  . In OBS, single
packets are assembled into units called bursts that are transmitted
in the optical domain. In general, prior signaling is carried out to
reserve network resources along a route, in order to build an all-
optical path. After a short time interval, called the offset time, the
burst is sent to the destination in the all-optical domain.
Optical burst switching has as its main advantage high network
utilization, since resources are reserved only if there is traffic
demand and they are released as soon as the burst has passed.
Also, there is low signaling latency due to the lack of
confirmation (in most signaling schemes) and a simpler
implementation compared to optical packet switching, since the
switching performance required for handling bursts is lower than
Grid networks based on lambda switching present some
disadvantages, as stated in . A dedicated path to every request
in a grid environment would lead to very high costs. Requests in a
grid are very unpredictable due to the high number of users;
therefore networks based on lambda switching are not suited for
these cases. Also, there are situations where the duration of a path
reservation is much smaller than the path setup time.
Optical burst switching is been considered as a candidate for grid
computing for many factors : Application jobs can be mapped
directly to optical bursts due to the variable granularity of the
bursts which allow different traffic profiles. The separation
between control data and application data allows an all-optical
transmission of the data bursts without signal conversion from the
optical to the electronic domain. Also, since the control packets
(BCP – Burst Control Packet) in OBS networks are processed
electronically, new features can be added for grid computing, like
3. ARCHITECTURE FOR CONSTRAINT-
BASED LSP SELECTION IN GRID OBS
3.1 GOBS Connection Server
To allow the selection of a route that attends the requirements of a
grid application, it is necessary that some resource information be
made available to the control and management planes. We assume
that this information is stored in components responsible for
determining the best path that matches the application’s
requirements. After that, GMPLS routing and signaling protocols
make effective resource reservations for the application.
The component that stores grid resource information and
calculates deterministic paths for an application will be called a
GOBS Connection Server. Through queries made by the grid
user/application, the GOBS Connection Server verifies which
possible routes are best suited to a specific request.
There are basically two types of resources in the context of this
study. A computing resource is a usually a grid node that process
and/or stores grid jobs. A network resource is an optical switch.
From now we use grid node to refer to computing resources and
network node for network resources.
When a node is inserted in the grid it needs to be registered at a
specific GOBS Connection Server, which will store the
information related to that node. This information refers to the
type of the node (grid or network), processing and storage
capacity, current blocking and class of service. The structure
suggested for the GOBS Connection Server is presented below
and is exemplified in Table 1. Figure 1 illustrates the GOBS
Connection Server in a typical context.
• Type of node (type): Defines if a node is a grid node or a
• Available processing capacity (processing): The current
processing capacity of a node measured in billions of
floating point operations per second (GFLOPs).
Applicable to grid nodes only.
• Available storage capacity (storage): The current storage
capacity of a node (in megabytes). Applicable to grid
• Blocking Probability (blocking): Applicable to network
nodes only. This parameter is obtained on-line based on
the number of blocked requests divided by the total
number of requests.
• Class of service (qos_class): Needed for service
differentiation. Applicable to network nodes only.
Table 1. Example of an entry in a GOBS Connection Server
Processing Storage Blocking
0 grid 2 1024 - -
1 network -
- 0.001 0
2 grid 5 512 - -
3 network - - 0.03 2
Figure 1. GOBS Connection Server
3.2 Route Selection
Route selection in this context is actually a multi-objective search
problem, that is, starting from a space composed by grid nodes,
which are connected to other nodes, a search for a path is
performed based on all application constraints.
This selection must consider parameters like the blocking
probability currently experienced by bursts in the nodes and the
processing and storage capacities available at a grid node.
Since in a grid there is no determination of which node is going to
be responsible for handling the task, we consider an anycast
approach  for resource reservation. The destination is fixed at
the moment a path is calculated in response to a request to the
GOBS Connection Server. After this, GMPLS explicit signaling
reserves all the resources along the path.
A search algorithm to allow route selection for an application will
now be described. The objective of this algorithm is to return an
explicit route to the OBS edge node that requested the path. This
response contains a route that matches the task demands and will
be used as input information for GMPLS signaling protocols.
Next, there is a high-level description of the search algorithm.
1: Determine the source node, mark source node as checked.
2: Starting from the source, check the next reachable nodes.
Include the source in the explicit route vector. Analyze the first
3: Check the type of the node.
4: If it is a grid node and it has not been checked:
5: Check if there is available processing/storage
capacity and mark as checked. In the case that the
request can be matched include this node in the explicit
6: If it is a network node and it has not been checked:
7: Check if the service levels (blocking) are within the
defined specifications and mark as checked. In the case
that the request can be matched include the downstream
node of the link in the explicit route vector.
8: Continue with the breadth-first search (repeat from step 3)
until it reaches a grid edge node capable of handling the request
9: If no path can be found then use a random destination.
4.1 Simulation description
We conducted several simulation studies in order to evaluate the
impact of the proposed architecture in the overall performance of
a Grid OBS network. The simulation tool used was ns-2 (version
The topology used in the simulations is illustrated in Figure 2. The
nodes that generate grid jobs are numbered 0, 1 and 2. Nodes 20
to 27 are grid computing nodes. The remaining nodes (3 to 19) are
network nodes (optical switches). The links have a capacity of 10
Gbps and a propagation delay of 1 millisecond. Each link has 8
wavelengths. The total traffic load is divided equally for the three
We defined three service classes for this study. Class 0 has the
highest priority and class 2 has the lowest. Table 2 shows the
constraints associated with each class.
Figure 2. Topology used in simulations
Table 2. Constraints associated with each QoS class
The sizes of the jobs are exponentially distributed with a mean of
1.5 megabytes , the processing demand is defined as a
percentage of the total processing capacity also exponentially
distributed, with a mean of 20%. The grid nodes (20 to 27) have
total storage and processing capacities of 1 gigabyte and 15
GFLOPS, respectively. The inter-arrival time of jobs follows a
Poisson process. The offset time is 3 milliseconds.
We use an admission control mechanism similar to the dynamic
wavelength grouping proposed in  where a maximum number
of wavelengths are associated with each service class, but there is
no imposition of which wavelength is associated with a particular
class. We defined class 0 to have 5 wavelengths, class 1 to have 3
and class 2 to have only 1 wavelength.
4.2 Result analysis
Figure 3 shows the blocking probability for all service classes
using the proposed route discovery scheme compared to default
Figure 3. Blocking probability vs. Class of service
It should be observed that the GOBS Connection Server allows
for the reduction of the blocking probability in a general way for
all service classes. This is due to the rerouting of bursts to other
destinations, which allows for a reduction in the traffic load sent
to a node and, in consequence, results in a smaller loss rate. It is
important to point out that all the blocking values are within the
constraints defined for each service class.
We also conducted an analysis of the utilization of the grid nodes
by evaluating the utilization of processing capacity of a grid node
(20 to 27). Figures 4 and 5 illustrate the mean processor and
storage utilization for the grid nodes.
Figure 4. Mean processor utilization at grid nodes
It is possible to observe that our proposal can reduce the
processing load at node 23, since it distributes these jobs to
several grid nodes, illustrating the anycast routing feature of the
Figure 5. Mean storage utilization at grid nodes
For storage the load is better balanced between the nodes, due to
the high traffic load offered to the network, and the route
calculation has little influence in the results, except for nodes 21
and 27 which have an higher increase in their storage utilization.
The mean end-to-end delay of a burst is illustrated in Figure 6.
Figure 6. End-to-end delay
The best delay result is for default routing. The increase in the
delay is a consequence of several queries to the GOBS
Connection Server, but, as can be seen, the difference is not too
Figure 7 shows the proportion of requests that could not be
guaranteed as a result of a query to the GOBS Connection Server
which was not successful in finding an adequate path for the burst.
Figure 7. Proportion of unsuccessful requests
Results show that, in default routing, for traffic loads above 25 Download full-text
erlangs, the majority of requests cannot be served adequately by
the GOBS Connection Server. However, using the proposed LSP
calculation scheme, the number of unsuccessful requests is below
30% of the total for loads above 40 erlangs.
5. RELATED WORK
Some anycast algorithms for OBS grid networks are presented in
. These algorithms differ in the way they perform burst
destination assignment and burst deflection. A number of
heuristics is proposed to deal with these issues. Our proposal
attempts to extend this anycast concept to the core nodes in
addition to destination nodes since it takes in consideration
network resource information to make the path calculation.
In  the authors propose an architecture based on the existence
of active elements called active OBS routers and on a two-stage
signaling scheme. In the first phase there is an active burst
responsible for inform the constraints of the job to the active core
nodes and for resource discovery. Next, in the second phase, the
normal burst in sent towards the selected destination. Our
proposal has similar aspects since it has a two-stage signaling,
where the first phase is the query to the GOBS Connection Server
and the second phase is the explicit signaling in response to this
query. The main difference is that we use a centralized approach
for route discovery, but not for signaling. In this way, we try to
take advantage of a complete view of resources and a rapid
This paper presented a model for the establishment of constraint-
based connections for Grid OBS networks based on GMPLS. Our
proposal has the advantages of combining both centralized and
distributed features to establish paths with precision and with low
latency, as well as using existing traffic engineering protocols to
enable the anycast feature of the proposal.
The anycast feature of the proposed architecture is extended to
deal with network resources in addition to grid nodes making
route selection more precise, and is very important if there is a
need to offer performance guarantees to certain applications.
One limitation of this proposal may be the reliance on the
existence of sufficient resources to handle all the requests. Traffic
engineering attempts to overcome this limitation by forwarding
bursts to nodes (grid and network) that may not be in use at that
Future work will be focused on a distributed alternative for the
GOBS Connection Server, which will be important for inter-
domain routing, and in the evaluation of other resource discovery
schemes that carry out reservation together with discovery.
We are also interested in addressing the scalability issues of the
model and in including more metrics in the job’s constraints
The authors would like to thank CAPES (Brazil) for financial
 Foster, I. Grid: A New Infrastructure For 21st Century
Science, Physics Today, 2002.
 Travostino, Franco; Mambretti, Joe.; Kamous-Edwards, Gigi.
Grid Networks: Enabling Grids with Advanced
Communication Technology. John Wiley and Sons, 2006.
 Chen, Y. Qiao C. e Yu, X. Optical burst switching: a new
area in optical networking research, IEEE Network, vol.18,
pp.16-23, May 2004.
 Simeonidou, D.; Nejabati, R.; Ciulli, N. (Editors). Grid
Optical Burst Switched Networks (GOBS): Informational
Track: draft-ggf-ghpn-GOBS-1. Jan, 2006.
 Habib, I.W.; Qiang Song; Zhaoming Li; Rao, N.S.V.
Deployment of the GMPLS control plane for grid
applications in experimental high-performance networks
Communications Magazine, IEEE, Vol.44, Iss.3, March
2006, Pages: 65- 73.
 Mannie E. (Editor). Generalized Multi-Protocol Label
Switching (GMPLS) Architecture. RFC 3945. Outubro,
 Chen, Y. Qiao C.; Yu, X. Optical Burst Switching: a New
Area in Optical Networking Research. IEEE Network,
vol.18, p.16-23, Maio, 2004.
 Vokkarane, V. M.; Jue, J. P. Optical Burst Switched
Networks. Springer, 2005.
 De Leenheer, M. et al. An OBS-based Grid Architecture. In:
Workshop on High-Performance Global Grid Networks
(Globecom, 2004). Proceedings of the IEEE Global
Communications Conference, Dallas, TX, USA, Nov. 2004.
 De Leenheer, M. et al, Anycast Algorithms Supporting
Optical Burst Switched Grid Networks. Proc. International
Conference on Networking and Services (ICNS), Silicon
Valley, USA, July 2006.
 NS-2. The Network Simulator. URL:
http://www.isi.edu/nsnam/ns/. Acessed in: January, 2008.
 Chen, Y.; Tang, W.; Verma, P. K. Latency in Grid over
Optical Burst Switching with Heterogeneous Traffic. In:
Proceedings of High Performance Computing and
Communications (HPCC 2007): 334-345.
 Zhang, Q.; Vokkarane, V. M.; Jue, J. P. e Chen, B. (2004).
Absolute QoS Differentiation in Optical Burst-Switched
Networks. IEEE Journal on Selected Areas in
Communications (JSAC), 22(9):2062-2071.
 Nejabati, R et.al, Programmable Optical Burst Switched
Network: A Novel Infrastructure for Grid, 5th IEEE/ACM
International Symposium on Cluster Computing and the
Grid, CCGrid 2005, Cardiff, UK, 9 - 12 May 2005.