Work in Progress: Harmony - Advance
Reservations in Heterogeneous Multi-domain
Environments
Alexander Willner1, Christoph Barz1, Joan Antoni Garcia Espin2, Jordi Ferrer
Riera2, Sergi Figuerola2, and Peter Martini1
1 University of Bonn, Institute of Computer Science IV, Römerstr. 164, 53111 Bonn,
Germany, {willner,barz,martini}@cs.uni-bonn.de
2 i2CAT Foundation, c/ Gran Capità 2-4, Edifici Nexus I, 2a planta, desp. 203, 08034
Barcelona, Spain,
{joan.antoni.garcia,jordi.ferrer,sergi.figuerola}@i2cat.net
Abstract. Grid computing aims at offering standardized access to het-
erogeneous and distributed resources for scientific communities. In order
to ensure certain quality of service requirements, the interconnecting net-
works have also be considered as Grid resources and must be taken into
account for the co-scheduling process. However, most current systems do
not support co-allocation of heterogeneous network resource provision-
ing systems and malleable advance reservations in multi-domain, multi-
technology, and multi-vendor environments. Our approach, called Har-
mony, provides a functional service plane to unify the underlying net-
work management systems and supports advanced reservation capabili-
ties to utilize the available capacity and network resources in an efficient
manner. The developed prototype has been demonstrated on numerous
conferences and a preliminary performance evaluation of the current im-
plementation is given.
Key words: Advance Reservations, BoD, GMPLS, MPLS, QoS
1 Introduction
Since its first demonstration in the context of the I-WAY [2] project more than
ten years ago, the concepts of Grid computing [3] have become increasingly
popular in the field of high performance computing. The vision is to virtualize
any kind of resource and to combine them to a single abstract entity that is as
simple to use as the power grid. One important research area in this context is the
co-allocation of these different resources. In particular, the underlying network
has to be considered as a first-class Grid resource that must be managed to
ensure certain quality of service (QoS ) requirements.
 Research for this paper was partially financed by the EU in the IST-034115 project
PHOSPHORUS [1]. We thank our project partners for their contributions and their
collaboration to this research work.

2 A. Willner, C. Barz, J. A. Garcia, J. Ferrer, S. Figuerola, P. Martini
As shown by Foster et al. [4] resource co-allocation in Grid computing envi-
ronments should be done automatically and transparently. In order to integrate
the network into this scheduling process, mainly two issues have to be faced.
Meanwhile the infrastructure has to support different QoS features, such as a
minimum bandwidth or maximum delay on a specified path, an agreement based
resource management is needed to reserve resources immediately or in advance.
In 2006, Travostino et al. [5] gave an overview of the relevance of network
research in Grid environments. In a nutshell, there have been two different ap-
proaches. Either low-level provisioning mechanisms for IP networks were used
such as Integrated Services (IntServ/RSVP) [6] and Differentiated Services (Diff-
Serv) [7]. Or, in order to meet the very high bandwidth demands of some Grid
applications, recent projects have started to use circuit-switched optical net-
works. However, shortcomings of the existing approaches are mostly their lim-
ited applicability in real environments or missing support of advanced scheduling
algorithms.
In this paper, we assume a set of independent administrative domains that
run their local network resource provisioning systems (NRPSs). These domains
might use different network technologies and are interconnected by static pre-
defined links. In our approach, we use an abstract service plane to provide an
end-to-end service and allow to co-allocate these networks in conjunction with
other Grid resources. The primary goal of this work is to provide a system that
reuses existing developments already running as NRPSs, and to allow indepen-
dent administrative domains to be integrated into the Grid resource scheduling
process, and therefore support advanced mechanisms to utilize the transport
network efficiently.
The key contribution of this paper is to provide a proof-of-concept and to
deliver measurements of a real implementation that builds an abstraction layer
for various existing signalling protocols. Overall, we believe this paper helps to
construct a foundation for further network research and developments in the field
of resource co-allocation in multi-domain, multi-technology, and multi-vendor
Grid environments.
The remainder of the paper is structured as follows. We give an overview of
related work in the context of dynamic circuit provisioning in Section 2. In Sec-
tion 3, we state our assumptions and the proposed architecture and we introduce
the challenges that are given in co-allocated multi-domain advance reservation
situations. Section 4 contains implementation details of our solution. In the sub-
sequent Section 5 a preliminary performance evaluation of our implementation
is given. Finally, we close with some conclusions and describe future work in
Section 6.
2 Related Work
Grid applications typically need to allocate and reserve multiple types of re-
sources, such as computation, data, instrumentation, and networks. The co-
allocation problem for computational Grids has been defined by Czajkowski [8]

Work in Progress: Harmony - Multi-domain Advance Reservations 3
1999. In the same year, based on the techniques and concepts of the Globus
Resource Management Architecture (GRMA) [9], a Distributed Resource Man-
agement Architecture (GARA) [10, 4] that used a Globus Resource Allocation
Manager (GRAM ) job scheduler to co-allocate the network with other resources
in advance was proposed. By then, the mechanisms used for network provisioning
were IntServ [6] and DiffServ [7].
While GARA is popular among the Grid community as a general purpose
platform allowing reservations of numerous resources it is not specialized for net-
works. Its API and Resource Specification Language (RSL) do not take network
specific attributes into account. Therefore, the Network Resource Scheduling En-
tity (NRSE ) [11] was introduced in the Grid Resource Scheduling (GRS ) project
in 2002. Still IntServ and DiffServ mechanisms were used to deliver guaranteed
throughput over packet-based networks.
Unfortunately, bulk data transfer oriented Grid computing often requires
guaranteed minimum bandwidth and minimized packet loss, which are not eas-
ily achievable in packet switching networks. Moving Terabytes or Petabytes of
data among multiple sites require dedicated optical networks, e.g., based on
wavelength switching, that can provide guaranteed bandwidth and performance
in terms of low bit error rates.
In 2005 the Exploitation of Switched Light paths for eScience Applications
(ESLEA) project had started. It demonstrated the usefulness of circuit-switches
networks for different application areas. The ESLEA Control Plane Software
(CPS ) was implemented as a modification of the afore mentioned NRSE and
was integrated into the EGEE Bandwidth Allocation and Reservation (BAR)
architecture. At the same time the DARPA DWDM-RAM project addressed
similar issues and a Network Resource Scheduling (NRS ) service was developed
to enable the efficient use of optical networks as a primary Grid resource.
Associated projects Harmony is based on existing solutions for intra-domain
path provisioning and scheduling, developed within associated projects. In total,
four different NRPSs are used to administer the underlying network: ARGIA [12]
(former UCLP – User Controlled Light path Provisioning); Nortels proof-of-con-
cept middleware called Dynamic Resource Allocation Controller (DRAC ) [13];
the MPLS [14] based Allocation and Reservation of Grid-enabled Optical Net-
works (ARGON ) [15] system that was developed within the German research
project VIOLA1; and also a thin adaption layer for the GMPLS control plane.
Related projects In fact, other related projects aim at similar challenges, al-
though from a different perspective, while the Phosphorus view is from the user
to the network: Originating from the GÉANT22 project, the Automated Band-
width Allocation across Heterogeneous Networks (AutoBAHN ) system; as an
achievement of the DANTE-Internet2-CANARIE-ESnet collaboration (DICE ),
1 http://www.viola-testbed.de
2 http://www.geant2.net/

4 A. Willner, C. Barz, J. A. Garcia, J. Ferrer, S. Figuerola, P. Martini
an inter-domain control plane based on OSCARS was developed where an inter-
domain controller (IDC ) communicate in a decentralized way to provision end-
to-end multi-domain network paths; the G-lambda project that develops an in-
terface between Grid resource management systems and network resource man-
agement systems that also support advance reservations; the GMPLS based En-
LIGHTened Computing project that focuses on dynamic optical light-paths be-
tween supercomputing sites which are created and torn down in advance or on
demand based upon application needs; the Dynamic Resource Allocation in GM-
PLS Optical Networks (DRAGON) project that aims at both the dynamic intra-
and inter-domain provisioning of packet and circuit switched network resources
in response to user requests for high-performance e-Science applications; and
finally, the Grid-enabled GMPLS (G2MPLS) Network Control plane, as a en-
hancement of the ASON/GMPLS Control Plane architecture that implements
the concept of Grid Network Services (GNS ).
3 Assumptions, Limitations and Architecture
Consider a network consisting of interconnected, independent domains that are
used as transit systems or that contain different kinds of resources. A domain in
this context is itself again a high performance (optical) network, with a control
plane above and a network resource management system (NRPS ) on top. Each
level of this architecture performs different roles and tasks in order to allow
creating on demand data plane paths for several users along a single domain.
Interconnection with third party domains has to be proposed, agreed and set up
by all involved parties. This leads to high delay in inter-domain path provisioning,
bureaucratic overhead and hard scheduling of resource utilisation (considering
both intra- and inter-domain resources) which most of the times discourage end
users.
The problem we consider here in detail is how to find and reserve an-end-to
end path between two or more resources on demand with certain QoS require-
ments, while supporting heterogeneous network technologies.As a consideration,
each domain must not reveal its internal topology and the network resources
should be utilized to capacity.
General Architecture The proposed architecture extends the afore mentioned
hierarchy by one additional layer. The network service plane (NSP) consists
of at least one inter-domain broker (IDB) and one NRPS Adapter (HNA) for
each subjacent domain. They both share the same Harmony Service Interface
(HSI ) and allow to exchange abstract topology information and to administer
network paths. The architecture has started with a centralized design (Fig. 1a)
and evolved to an hierarchical (Fig. 1b) and distributed (Fig. 1c) system.
Topology Knowledge In particular, for reasons of confidentiality the internal
topology information of a domain might not be opened to the service layer. This
is the reason why the knowledge of the global topology is restricted to a set

Work in Progress: Harmony - Multi-domain Advance Reservations 5
n s p
a rg o n Ad a p te r g mp l s Ad a p te r u c l p Ad a p te r d ra c Ad a p te r
a rg o n g mp l s u c l p d ra c
(a) centralized
n s p
n s p a r go A d n s p a p d p t e
d em o n lu d p ct e m p A s lu d p ct e  A p lu d p ct e u ed  lu d p ct e
d em o n m p A s  A p u ed 
(b) hierarchical
n sp a r g o A d t
n sp a r g o A d e
r ml p o uc r d  mt
r ml p o uc r d  me
 m g o A d e
r ml p o uc r d  m m g o A d t
se  r g o A d t
se  r g o A d e
 m   a d uc r d  mt
 m   a d uc r d  me  a d uc r d  mt
 a d uc r d  me
r ml p o t
r ml p o e
r ml p o  m   a d t
 m   a d e  a d t
 a d e
(c) distributed
Fig. 1. The service plane architectures and the operating modes implemented.
White circles are IDBs, grey circles are adapters, and grey boxes are administra-
tive domains with their corresponding NRPS.
of basic information based on three main elements: the endpoint, the link, the
domain itself, and the border points.
Based on the Transport Network assigned address (TNA) [16] endpoints in
the NSP are identified by strings with IPv4 syntax. Each network port receives
a unique TNA in the domain it is attached to and the domain itself serves
one or more TNA address ranges. Two border endpoints identify uniquely an
inter-domain link between two given domains.
Every domain exports its border endpoints connected to inter-domain links
and the inter-domain links themselves. Hence, a transport network controlled by
a single HSI capable system is seen as a cloud with a set of border endpoints.
If two or more domains are controlled by a single IDB, this new super-domain
will be seen as one cloud with a subset of the original border endpoints. Border
endpoints connected to the other domains under control of the same IDB are
kept as intra-domain endpoints and are not passed up.
Path Computation End-to-end path provisioning requires dynamic intra- and
inter-domain path computation. Since only an abstracted topology knowledge
is distributed, the intra-domain path computation is performed by each NRPS.
The inter-domain path computation is performed inside each IDB by using Di-
jkstra’s shortest path algorithm. Only border endpoints and inter-domain links
are considered.
A reservation itself contains one or more services and each service contains
one or more connections. Whereby each connection is a requested path between
one source TNA and one or more destination TNAs. Different fixed or malleable
QoS demands can be defined on each level.
The actual reservation is following a non-blocking tree two-phase commit
protocol scheme. After the reception of a reservation request, the path computer
starts looping over all known network resources in order to find feasible paths.
Upon the completion of this task, the availability of the resources within all

6 A. Willner, C. Barz, J. A. Garcia, J. Ferrer, S. Figuerola, P. Martini
involved domains in the calculated path is requested. In case the resources are
not available, they are pruned within the path computer session and the loop
will start again. Otherwise the resource will be reserved.
Malleable Multi-domain Network Resource Allocation As stated in [17]
and depicted in Fig. 2 reservations can be divided into two different types. First,
reservations can start straightaway upon receipt of the reservation (Fig. 2a). Sec-
ond (Fig. 2b), they can be scheduled with a specified starting time and a specified
duration (fixed, FAR, STSD), a specified starting time and a unspecified dura-
tion (STUD), an unspecified starting time and a specified duration (deferrable,
DAR, UTSD), and finally with an unspecified starting time and an unspecified
duration (malleable, MAR, UTUD).
time
bandwidth
b
a
(a) immediate reservations
time
bandwidth
c
d
a b
(b) advance reservations
Fig. 2. The different types of reservations. Fading areas represent unspecified
bandwidth or time constraints.
Of particular interest are the advance reservation requests with elastic start-
ing times and an unspecified duration. These malleable reservations can be used
by Grid middleware applications for file transfers. They allow the largest flexi-
bility to schedule resources and to utilize the full network capacity.
When a malleable reservation request is received, the IDB starts looping over
distinct inter-domain paths, with feasible start-times and bandwidths in order
to find a set of available resources to fulfil the request. First of all, the possible
paths are calculated. After that, the path computation algorithm starts looping
over all the obtained paths. Inside the loop, the feasible bandwidths of the end-
points involved in the connection are retrieved and the maximum bandwidth in
the range of the bandwidth provided in the request is chosen. This bandwidth is
adjusted in all endpoints to the given path according to the technology-induced
granularity. In a final step, the algorithm checks for the availability of the network
resources. In case any NRPS returns non-available result, the IDB begins to
adapt the start time and bandwidth.
Formal Definitions We define the data plane as digraph Gd = (D;L). The el-
ements of set D are called domain, and the elements of L are ordered pairs of ver-
tices, called inter-domain link with (ud; vd; idd) 2 L  DD ID. An inter do-
main path inGd between two domains dsource and ddest is an adjacent sequence of

Work in Progress: Harmony - Multi-domain Advance Reservations 7
inter domain links pd(dsource; ddest) = (ud0 ; vd0 ; idd0); : : : ; (udm1 ; vdm1 ; iddm1),
with ud0 = dsource, vdm1 = ddest, m 2 N0 and len(pd) = m. The number of
parallel inter-domain links between two domains can be defined as n(ud; vd) =
jf(udi ; vdi ; iddi) 2 Ljud = udi ^ vd = vdigj. The successor domains of a domain
referring to the data plane are defined as a function s(ud) = fwdj9(ud; wd; idd) 2
Lg. Let pid(dsource; ddest) = (udi ; vdi ; iddi) be the ith link of a path. Then the ad-
jacent domains of a path pd are defined as a(pd) = fdsourceg[0i<len(pd)fvdi jp
i
dg.
The service plane is described as the digraph Gs = (H;C) with H a set
whose elements are called Harmony Service Interface entity (HSI entity), and C
a set of ordered pairs of vertices, called HSI interconnection with (us; vs) 2 C.
In the centralized or hierarchical model these HSI interconnections also reflect
the hierarchy. A path in Gs in denoted as ps = (us0 ; vs0); : : : ; (usm1 ; vsm1),
with m 2 N0 and len(ps) = m. The child of an HSI adapter is defined as
sus = fwsj(us; ws) 2 Cg. This means, that the HSI entity us delegates the
reservation task to the set s(us). Thus, we have a tree structure of HSI entities.
In the distributed model we have a mesh of HSI entities which may be themselves
the root of a HSI entity tree.
Let the set of Domains for which an HSI entity h is responsible be r(h 2
H) 2 D and R(Hi  H) = fr(h)jh 2 Hig. Thus, the set of responsible HSI
entities of a data path is R(a(pd)).
4 Implementation
The communication between the entities is done by using SOAP messages and
the Apache Muse framework. To support workflows and resource co-allocation,
the API was derived from existing interfaces used to abstract and map network
resources with most limited knowledge in intra-domain Grid environments and
was modified to cover aspects of multi-domain conditions [18].
The underlying testbed (Fig. 3a and Fig. 3b) has a high similarity to real net-
work environments in the National Research and Education Networks (NRENs).
Due to its heterogeneity and variety of administration actors, testing of new
developments in the service plane is more controllable when done in a virtual
testbed (Fig. 4a and Fig. 4b).
The data plane inter-connections within the testbed use dedicated light-paths
from either GÉANT2 or GLIF3 infrastructure. The switching equipment in the
local testbeds is composed among others by the following systems: Alcatel-Lucent
1678, 1850, and 7750; Calient DiamondWave FiberConnect; Cisco Catalyst 6509,
3750; Nortel Optera 5200, OME 6500, and HDXc; and Riverstone 15008. Fur-
thermore, hardware used to run the NSP was compiled by standard personal
computer components and virtual machines. The main IDB was running on an
Intel Quad Core Xeon with 2.80 GHz, 1 GByte RAM, and Fedora Core release
6.
3 http://www.glif.is/

8 A. Willner, C. Barz, J. A. Garcia, J. Ferrer, S. Figuerola, P. Martini
(a) real data plane (b) real service plane
Fig. 3. The data and service plane testbed architecture. Whereby grey boxes are
administrative domains with their corresponding NRPS and arrows are precon-
figured VLANs between them. Grey circles are domains controlled by a single
NRPS and white circles represent IDBs.
(a) virtual data plane (b) virtual service plane
Fig. 4. The service and the emulated data plane architecture within the virtual
testbed. Whereby grey boxes are administrative domains with their correspond-
ing NRPS and arrows represent emulated VLANs between them. Grey circles
are domains controlled by a single NRPS and grey circles represent IDBs.

Work in Progress: Harmony - Multi-domain Advance Reservations 9
5 Evaluation
Formal Performance Analysis First, we consider the centralised and hierar-
chical model for the formal performance analysis. We show, how the processing
times of the different domains can be combined for the overall service time of
a single path request. Based on this definition, we extend the approach to a
distributed model. Subsequently, we show a worst case calculation of the neces-
sary configuration steps if there are several alternative data links between the
domains and reservation attempts fails.
The time a single HSI entity h 2 R(a(pd)) needs to process a request inter-
nally is denoted as th and delegates the reservation to its child HSI entities. If
we assume that all children of h are requested consecutively, we can define ttotalh
of an HSI entity recursively as:
ttotalh =
(
th if js(h)j = 0,
th +
P
hi2s(h)\R(a(pd))
ttotalhi if js(h)j)  1.
(1)
If the communication with the children is parallelized, we can define thtotal
as:
ttotalh =
(
th if js(h)j = 0,
th + maxhi2s(h)\R(a(pd)) t
total
hi if js(h)j)  1.
(2)
In case of using the distributed model for the service plane is used there is a
mesh of HSI entities on top level Htop 2 H. Each top level HSI entity htop 2 Htop
can itself be the root of a hierarchical HSI entity structure.
Thus, we get for path request pd for a distributed service plane operating in
a consecutive mode:
ttotal =
X
hi2Htop\R(a(pd))
ttotalhi (3)
For a path request for pd for a distributed service plane operating in a parallel
mode:
ttotal = max
hi2Htop\R(a(pd))
ttotalhi (4)
with ttotalhi to be evaluated with Eq. 3 or Eq. 2.
If there are several links between at least two adjacent domains of a path
pd(dsource; ddest and the reservation request fails in using one of these redundant
links, there might still be a chance in reserving the data path via one of the
alternative links. In a consecutive model, where a local reconfiguration is possible
as a crankback mechanism, the number of configuration steps c of a path p is
limited by:
O(c) = len(p) +
X
0ilen(p)1
n(ui; vi) 1 (5)

10 A. Willner, C. Barz, J. A. Garcia, J. Ferrer, S. Figuerola, P. Martini
Here, the first summand represents the successful configurations and the second
summand the reconfigurations because of a failed reservation request at one of
the HSI entities.
Measurements To emphasize and support the preceding analysis, measure-
ments from the running prototype are given. As shown in Fig. 5a the mean
response time of each NRPS adapter (th) varies from 200 ms to 410 ms and has
a noticeably large number of outliers. The dummy adapter delay points out a 10
ms communication and processing overhead for each request. As seen in Fig. 5b
each additional hierarchy level increases the total response time (ttotalh ) by ap-
proximately 500 ms. Furthermore, Fig. 6a and Fig. 6b indicate that the current
prototype’s response time increases by arithmetic progression with reference to
the amount of incoming requests. It also shows clearly a performance threshold
at about 39 requests per second. After that the success rate drops dramatically
and the request duration almost triples.
lllllll
ll
l
lll
l
l
l
l
l
l
ll
ll
l
ll
l
l
ll
l
l
l
l
ll
l
ll
llllll
l
l
ll
Dummy Thin/GMPLS ARGIA Argon
0
200
400
600
800
1000
1200
1400
involved adapter
durat
ion [ms
]
(a) direct calls
l
l
lllll
lllll
l
llll
l
l
l
ll
l
l
l
l
l
ll
l
l
l
l
ll
l
l
ll
l
llll
l
l
l
l
l
l
l
l
l
l
l
l
ll
l
l
l
ll
0 1 2
500
1000
1500
2000
2500
involved broker [#]
durat
ion [ms
]
(b) two levels of hierarchy
Fig. 5. Processing of a reservation request within a single IDB (500 repetitions).
(a) depicts the response time for each Adapter and (b) shows the delay that is
added by every IDB hierarchy level.
6 Conclusion and Future Work
We have shown that under the given assumptions and limitations the proposed
architecture allows for multi-domain, multi-technology, multi-vendor network
reservation and co-scheduling. We’ve tested it in a real scenario under the Phos-
phorus Europe wide testbed. Besides this proof-of-concept the crucial result is

Work in Progress: Harmony - Multi-domain Advance Reservations 11
(a) duration of each successful reservation (b) success rate
Fig. 6. Processing of n reservation request per second within the Service Plane.
(a) depicts the duration of each successful request and (b) shows the success
rate.
that even with a very limited knowledge of topology information and without an
homogeneous control plane, complex malleable advance reservations are feasible
and realizable.
Based on these results different network operators such as National Research
and Education Networks (NRENs) would be able to establish bilateral agree-
ments for end-to-end provisioning services without changing technologies or ex-
posing confidential data.
Besides the implemented prototype that certainly is valuable to demonstrate
the main functionality within a testbed with real hardware and users, simula-
tions would provide the possibility to evaluate alternative algorithms in a more
sophisticated manner. We expect to have first results using the discrete event
simulation package SimJava [19] shortly. Moreover, we are currently implement-
ing more progressive algorithms for malleable reservations in order to reduce the
number of availability requests and increase the average network utilization.
Long-term objectives mainly focus on enhanced network resiliency capabili-
ties. This includes as well the computation of disjoint paths as the delegation of
responsibilities, path monitoring, and automatic re-routing and -connection.
References
1. Figuerola, S., Ciulli, N., de Leenheer, M., Demchenko, Y., Ziegler, W., Binczewski,
A., et al.: PHOSPHORUS: single-step on-demand services across multi-domain
networks for e-science. In: Network Architectures, Management, and Applications
V. Edited by Wang, Jianli; Chang, Gee-Kung; Itaya, Yoshio; Zech, Herwig. Pro-
ceedings of the SPIE. Volume 6784. (2007) 67842X
2. Foster, I., Geisler, J., Nickless, B., Smith, W., Tuecke, S.: Software Infrastructure
for the I-WAY High-Performance Distributed Computing Experiment. In: HPDC.
Volume 96. (1996) 562–571

12 A. Willner, C. Barz, J. A. Garcia, J. Ferrer, S. Figuerola, P. Martini
3. Foster, I., Kesselman, C.: The Grid: Blueprint for a New Computing Infrastructure.
Morgan Kaufmann (2004)
4. Foster, I., Jennings, N., Kesselman, C.: Brain Meets Brawn: Why Grid and Agents
Need Each Other. In: International Conference on Autonomous Agents: Proceed-
ings of the Third International Joint Conference on Autonomous Agents and Mul-
tiagent Systems. Volume 1. (2004) 8–15
5. Travostino, F., Mambretti, J., Karmous-Edwards, G.: Grid Networks: Enabling
Grids with Advanced Communication Technology. John Wiley & Sons (2006)
6. Braden, R., Clark, D., Shenker, S., et al.: Integrated Services in the Internet
Architecture: an Overview. IETF RFC 1633 (Proposed Standard) (1994)
7. Blake, S., Black, D., Carlson, M., Davies, E., Wang, Z., Weiss, W.: An Architecture
for Differentiated Service. IETF RFC 2475 (Proposed Standard) (1998)
8. Czajkowski, K., Foster, I., Kesselman, C.: Resource Co-Allocation in Computa-
tional Grids. In: Proceedings of the Eighth IEEE International Symposium on
High Performance Distributed Computing (HPDC-8). (1999) 219–228
9. Czajkowski, K., Foster, I., Karonis, N., Kesselman, C., Martin, S., Smith, W.,
Tuecke, S.: A Resource Management Architecture for Metacomputing Systems.
LECTURE NOTES IN COMPUTER SCIENCE (1998) 62–82
10. Foster, I., Kesselman, C., Lee, C., Lindell, B., Nahrstedt, K., Roy, A.: A dis-
tributed resource management architecture that supportsadvance reservations and
co-allocation. In: Quality of Service, 1999. IWQoS’99. 1999 Seventh International
Workshop on. (1999) 27–36
11. Bhatti, S., Sorensen, S., Clark, P., Crowcroft, J.: Network QoS for Grid Systems.
International Journal of High Performance Computing Applications 17 (2003) 219
12. Grasa, E., Junyent, G., Figuerola, S., Lopez, A., Savoie, M.: UCLPv2: a network
virtualization framework built on web services [web services in telecommunications,
part II]. Communications Magazine, IEEE 46 (2008) 126–134
13. Travostino, F., Keates, R., Lavian, T., Monga, I., Schofield, B.: Project DRAC:
creating an applications-aware network (2005)
14. Rosen, E., Viswanathan, A., Callon, R., et al.: Multiprotocol Label Switching
Architecture. IETF RFC 3031 (Proposed Standard) (2001)
15. Barz, C., Bornhauser, U., Martini, P., Pilz, M., de Waal, C., Willner, A.: ARGON:
Reservation in Grid-enabled Networks. In: Proceedings of the 1. DFN-Forum on
Communication Technologies. (2008)
16. Rajagopalan, B.: Documentation of IANA Assignments for Label Distribution Pro-
tocol (LDP), Resource ReSerVation Protocol (RSVP), and Resource ReSerVation
Protocol-Traffic Engineering (RSVP-TE) Extensions for Optical UNI Signaling.
IETF RFC 3476 (Informational) (2003)
17. Varvarigos, E., Sourlas, V., Christodoulopoulos, K.: Routing and scheduling con-
nections in networks that support advance reservations. Computer Networks (2008)
18. Andree, B., Hommes, F., Peeters, B., Recio, J., Sánchez, A., Schon, J., de Waal,
C., A.Willner: PHOSPHORUS Deliverable D1.1.: Requirements and specifications
of interfaces architecture for interoperability between NRPS, GMPLS, Middleware.
European IST project PHOSPHORUS (2007)
19. Howell, F., McNab, R.: SimJava: A Discrete Event Simulation Package For Java
With Applications In Computer Systems Modelling. In: Proceedings of the First
International Conference on Web-based Modelling and Simulation. (1998)