1. DFN-Forum 2008
Themenkreis II (Grid-Technologien)
ARGON: Reservation in Grid-enabled Networks
Vortragender
Markus Pilz
Universita¨t Bonn
Institut fu¨r Informatik IV
Ro¨merstr. 164
53117 Bonn
pilz@cs.uni-bonn.de
+49 228 73 4549
Co-Autoren (selbes Institut)
 Christoph Barz (barz@cs.uni-bonn.de)
 Uli Bornhauser (ub@cs.uni-bonn.de)
 Prof. Dr. Peter Martini (martini@cs.uni-bonn.de)
 Dr. Christian de Waal (dewaal@cs.uni-bonn.de)
 Alexander Willner (willner@cs.uni-bonn.de)

ARGON: Reservation in Grid-enabled Networks
Christoph Barz, Uli Bornhauser, Peter Martini,
Markus Pilz, Christian de Waal, Alexander Willner
Institute for Computer Science IV, University of Bonn, Germany
fbarz,ub,martini,pilz,dewaal,willnerg@cs.uni-bonn.de
Abstract: Grid computing offers heterogeneous and distributed resources to scientific
communities. Apparently, networks connecting these resources can also be consid-
ered as Grid resources. This paper presents ARGON, a system that integrates metro
and wide area networks into Grid environments by providing advance reservations
and guaranteed network services. Here, single-domain as well as multidomain net-
work environments are considered. A major objective is to support metaschedulers in
the planning of workflows for e-science applications with demanding network require-
ments.
1 Introduction
One objective of Grid computing [F+03] is to offer a standardized interface to heteroge-
neous resources such as computational clusters, data storage sites, and scientific instru-
ments. These resources are distributed and typically interconnected via the Internet. Jobs
within a Grid are often specified as workflows. As part of the workflow planning and exe-
cution, different types of network services can be used. This includes the transfer of data
as well as network connectivity for synchronous streaming and coupling of parallel jobs
between different Grid sites.
According to [Fer07], the analysis of local and remote data becomes more and more im-
portant in the field of e-science, medicine, engineering, and digital art. These data sets or
streams may be as large as several terabytes, and may be real-time or preprocessed. It is
important to guarantee that the data to be processed are delivered timely. It is envisioned
that requests for multiple Gbps need to be fulfilled in order to leverage these applications.
Since currently this requested quality of service (QoS) cannot be ensured in the Internet,
many Grid sites are additionally interconnected by high-speed networks. These networks
usually just provide manually arranged connections with a given QoS.
However, even in a high-speed network environment it might not be feasible to overpro-
vision the network links, and the available bandwidth must be shared and allocated in an
efficient manner. Furthermore, a static network setup restricts the selection of available
Grid resources. In addition to high bandwidth requirements, interactive and collaborative
environments may require specific round trip times, a low packet loss, and small jitter or a
combination of those.
In order to allow for a flexible assortment, a system is needed that allocates network paths

on demand and offers this service to higher layers. This system needs to align the re-
quirements of the Grid applications with the traffic engineering capabilities. As different
Grid resources may have to be coallocated with the network resources, it is also necessary
to make reservations in the network in advance. Usually, this coordination of different
Grid resources is done by a metascheduling service (MSS). As a consequence, a network
management system is an additional system to be queried by such a scheduler.
Both the demand of applications on the network and the above mentioned challenges lead
to the design and implementation of ARGON (Allocation and Reservation of Grid-enabled
Optical Networks). ARGON has the ability to abstract the network resources and makes
them available by providing a reservation interface. On the one hand, ARGON integrates
into an existing MSS [B+07a] that orchestrates and reserves all available Grid resources
including a single network domain. On the other hand, it may be used by an interdomain
system like the Network Service Plane (NSP) [F+07] to allow for a network allocation in
multidomain environments.
This paper is structured as follows. In Section 2, the related work in this field is presented.
We then proceed to present ARGON, our proposal for such a network reservation system
in Section 3. In the following Section 4 we extend the considered objective to interdomain
environments. Finally, future work is discussed and our conclusions are summarized in
Section 5.
2 Related Work
A starting point for ARGON (Allocation and Reservation of Grid-enabled Optical Net-
works) was the German research project VIOLA (Vertically Integrated Optical Testbed
for Large Applications in DFN) [VIO07]. The project established a vertically integrated
approach to combine scientific applications, Grid middleware, and advanced networking
equipment. In this section, we mainly focus on related projects and architectures that allow
for an automated configuration of network resources for Grid applications. While there is
no common terminology for such a system, we adopt the term Network Resource Man-
ager (NRM) following the terminology of the Grid community. Other projects or systems
might use terms related to network management.
Projects mainly focusing on intradomain configuration: The UCLP (User Controlled Light-
path Provisioning) system [UCL07] (the commercial Version is called ARGIA) is designed
to allow end users to provision and dynamically reconfigure end-to-end lightpaths in op-
tical networks across a single domain or multiple independently managed domains. UCLP
supports the delegation of the control of subnetworks to other users. It is designed to allow
for the creation of application-specific IP networks for high-end e-science and Grid appli-
cations. Nortel developed a proof-of-concept middleware called DRAC [T+05] (Dynamic
Resource Allocation Controller) that allows for an application initiated configuration of
network resources on an end-to-end basis. DRAC is able to coordinate optical and packet
switched networks on demand or based on reservations. The main goal of the G-lambda
project [T+06a] is to establish a standardized Web service interface between Grid resource

management systems and network resource management systems that also support ad-
vance reservations. The main focus of the project is on optical network technologies. The
EnLIGHTened Computing project [B+07b] focuses on dynamic optical lightpaths between
supercomputing sites which are created and torn down in advance or on demand based
upon application needs. A domain manager allocates network resources by dynamically
setting up and deleting dedicated circuits using GMPLS control plane signalling.
Projects mainly focusing on interdomain configuration: Originating from the GE´ANT2
project, the AutoBAHN (Automated Bandwidth Allocation across Heterogeneous Net-
works) architecture targets at the needs of a multidomain, multitechnology research and
communication community. Based on the InterDomain Manager (IDM) of JRA3 [C+06],
the AutoBAHN architecture defines an interdomain network reservation mechanism based
on a decentralized architecture for interdomain signalling. In addition, an interface to Do-
main Managers (DM) is defined. The DMs are responsible for the management and con-
figuration of the local domains. The DRAGON project (Dynamic Resource Allocation in
GMPLS Optical Networks) [T+06b] aims at both the dynamic intra- and interdomain pro-
visioning of packet and circuit switched network resources in response to user requests for
high-performance e-science applications. Its architecture is based on the GMPLS control
plane stack and allows for deterministic services on an interdomain basis and across het-
erogeneous network topologies. DRAGON uses a peer or augmented interdomain model
supporting the exchange of abstracted topology information for signalling and interdomain
path computation. Its model is similar to the IETF Path Computation Element Archi-
tecture [FVA06]. As an achievement of the DICE (DANTE-Internet2-CANARIE-ESnet)
collaboration [DIC07] of CANARIE, ESnet, GE´ANT2 and Internet2, a web-based inter-
domain control plane was developed where IDCs (InterDomain Controller) communicate
in a decentralized way to provision end-to-end multidomain network paths. Every IDC ad-
vertises an abstracted topology. As part of the admission procedure, an interdomain path
is computed and signalled. The internal configuration of the domains involved is explicitly
out of scope of the collaboration.
3 ARGON – An approach to intradomain network reservations
The following section describes the way ARGON factors network resources into a Grid
environment. Core components of and services provided by the NRM are presented. This
section concludes with a discussion on processing advance reservation requests.
3.1 ARGON’s Core Components
Authentication and authorization (AA) are a cornerstone in the context of Grid computing.
Every request processed by ARGON’s AA Management is obliged to contain correspond-
ing information about the requester. Immediately upon the receipt of a new request the
contained AA information is being analyzed. Successfully authenticated and authorized

requests are passed to the Request Handler.
The Request Handler dispatches messages depending on their type (cf. Section 3.2) to
corresponding components and handles replies which contain processing results and the
status of reservations.
The Resource Management keeps track of available and allocated network resources.
Note that while reservations are being made via the resource reservation system, spare
network capacity can still be used for best effort traffic if traffic shaping is supported by
the network components.
Using the information provided by the Resource Management, the Path Computer deter-
mines feasible resource allocations (cf. Section 3.3). While availability requests are only
processed by the path computer, allocated resources are reserved if a reservation request
is handled. After a successful allocation of resources, the Scheduler is informed about the
corresponding starting and ending time of the reservations.
Once a reservation is accepted, the Scheduler is used to trigger the network configuration.
This is necessary because network components used in today’s systems are not aware of
time and do not support scheduled configuration.
The Signalling Subsystem is triggered by the Scheduler and is responsible for the setup
and tear down of paths in conjunction with traffic shaping configurations in the network.
Considering the processing of an entire request, the configuration of the network com-
ponents can be a time consuming process: Command line interfaces still dominate the
configuration process of routers as comprehensive machine-machine interfaces are rare.
Currently, MPLS and GMPLS entities from different vendors are supported.
3.2 ARGON’s Supported Services
ARGON supports six different services to factor network resources into the Grid by sup-
porting advance reservations. These services and the corresponding interface support
a metascheduler in planning Grid application workflows with network specific require-
ments [B+07a].
A Reservation Request is used to make an obligatory reservation of network resources
in advance. The request itself defines a workflow containing a hierarchy of three levels.
The top level, denoted as reservation, aggregates several different services to a single
workflow. Such a workflow is accepted or rejected as a whole by the reservation system.
A service in turn comprises multiple connections between different endpoints specifying
the same service type with the same time constraints. Connection parameters comprise a
set of constraints like bandwidth or delay constraints as well as binding information (cf.
Binding Request).
The Availability Request specifies the same parameter as a Reservation Request. This
allows a user or metascheduler to check whether resources are currently available. If a
requested workflow is unfeasible, the reservation system replies with the first starting time
at which the requested workflow is feasible. This information facilitates the planning of

alternative configurations for temporary unfeasible grid jobs.
A Cancel Request is used to revoke an accepted reservation request. Reservations which
are not yet in their usage phase are trivial to cancel, as only resource allocation information
has to be adjusted. In contrast, cancelling a reservation during its usage phase requires a
reconfiguration of routers in the network.
The Modification Request is used to change accepted reservations. The processing of
a modification is handled in such a way that resources reserved by the original request
cannot be allocated by other users. So, if a modification fails, the original reservation
remains.
A reservation made in advance can be specified in such a way that either the configuration
of network devices is triggered automatically by the Scheduler entity or explicitly triggered
by an additional request. The Activation Request is used in the latter case to initiate the
signalling process. Hence, allocated but unused resources might be temporarily be used
by best effort traffic or other reservations.
The Binding Request can be used to specify additional information necessary for pro-
visioning purposes. As an example, a binding is required if multiple cluster nodes are
connected to the same endpoint (e.g. a router) of a connection and only a subset of the
connected nodes participate. The parameters define the traffic to be mapped on a tunnel at
the endpoint by using IP address and port information. A separate request type is required
for specifying this information, because these parameters may not be available at the point
in time when the reservation is made. This allows a reservation to be made although the
involved cluster nodes are still unknown.
3.3 Processing Advance Reservations for Network Resources
Advance reservation requests contain time as well as resource related parameters that have
to be mapped on available network resources. The task of the path computer is to compute
an optimal schedule for a single request as well as to reschedule accepted requests to
optimize the current resource usage. After introducing the resource management utilized
in ARGON, the processing of advance reservations is discussed.
3.3.1 Resource Information Management
Information about the allocation of network resources for previously accepted requests is
required to determine the feasibility of a new request. The resource information manage-
ment keeps track of residual capacities (or usage profiles) in the network topology, i.e. the
topological structure as well as allocated resources are regarded with respect to time. This
information is temporally limited by a so-called book-ahead interval. The timeslot based
management of allocated resources is an established way [GO00] to manage this informa-
tion as it allows to decouple the computational complexity of the admission decision from
the number of already accepted requests.

ARGON uses an enhanced timeslot model [BBMP08]: Timeslots have a dynamic length
with a fixed granularity. In brief, for every new reservation, existing intervals may be split
at the start and the end time of a requested connection. So, every time a connection of
a reservation is accepted, no more than two new timeslots are created in addition to the
already existing ones. This results in a timeline being segmented in at most 2k + 1 non-
overlapping subintervals, where k is the number of accepted requests. It should be noted
that even if the number of timeslots depends on the number of reservations, an upper bound
is given by the length of the book-ahead and the granularity.
3.3.2 Fixed and Deferrable Advance Reservations
A basic form of an Advance Reservation (AR) request is defined as follows: The request is
received at tarrival, is admitted and starts at tstart. Furthermore, the usage phase (duration)
is limited by tend. This life cycle is depicted on the left hand side in Figure 1.
A fixed advance reservation request for a single connection is defined as tuple (tstart; tend;
s; d; C) where tstart < tend. The reservation starts at tstart and ends at tend. The endpoints
of the connections are specified by s and d. C represents additional resource constraints
which are usually the required capacity and delay constraints. The resource allocation
information of the timeslots involved in the request is accessed to determine whether a
request is feasible or not. The time complexity of the path selection (or path routing) is de-
pendent on the constraint set. If only one connection with link constraints (e.g. minimum
capacity) is requested, a shortest path algorithm with a polynomial running time identi-
fies a candidate using a constrained topology. If multiple paths are requested at the same
time or other path constraints (e.g. loss rate, delay) are included, the complexity of the
path selection process can increase to a super-polynomial time complexity. In these cases,
heuristics with potentially suboptimal results can be applied to keep the processing time
low, e.g. multiple paths within a request are processed independently one after another.
The separation of the points in time at which a reservation request is received and at which
it is supposed to begin results in a new degree of freedom for the reservation system.
It may not be necessary to reply an advance reservation request directly. However, it is
assumed that the reservation system operates like an online system, meaning that the ne-
gotiation phase is as short as possible. This is very important for coallocating resources,
as following reservations of non-network resources may depend on the admission control
decision made by the NRM. Nevertheless, if the resource usage in a certain time inter-
val exceeds a threshold, a rescheduling process can be started to optimize the currently
accepted reservations. The time complexity of this process is again dependent on the ac-
cepted reservations and corresponding constraints. Heuristics for this offline optimization
problem are currently under development.
An additional type is a deferrable advance reservation which has a certain degree of free-
dom in the time domain. In particular, time related parameters define a range of possible
values to establish the reservation. The life cycle of a deferrable advance reservation is
given on the right hand side in Figure 1 and defined as tuple (trelease; tdeadline;
t; s; d; C)
where trelease +
t < tdeadline. The reservation can start at trelease and must end before
tdeadline. The length of the usage phase is specified by the duration
t > 0. Compared

potential usage phase
tarrival tstart
negotiationphase usage phasependingphase
tendtrelease tdeadlinetarrival tstart
negotiationphase usage phasependingphase
tend
Figure 1: Life cycle of a fixed advance reservation (left) and a deferrable advance reservation (right)
to a fixed advance reservation, the parameters tstart and tend (
t = tend tstart) can be
determined by the NRM. An option is to specify the minimum and maximum length of
a usage phase as an interval. Again, various strategies can be introduced to handle this
parameter range, e.g. to reduce the usage phase down to the minimum value, if this allows
for the admission of additional requests.
3.3.3 Malleable Advance Reservation
A specification of the exact transmission rate can be omitted when a fixed amount of data
has to be transmitted. Only general capabilities of the sender and the receiver, such as the
maximal transfer rate and timing constraints for the transmission like a fixed deadline or an
earliest starting time (release time), have to be regarded. By joining the time and resource
constraints, the reservation system can find the most efficient solution for the requested
transmission. This kind of reservation is denoted as malleable advance reservation [Bur04]
(or advance cumulative reservation [GO00]). A motivating example regarding efficiency
is to fill gaps between allocated resources which are caused by accepted reservations.
A malleable reservation request is defined as tuple (trelease; tdeadline; s; d; S; C) where
trelease < tdeadline. The endpoints of the connections are specified by s and d. S de-
termines the data size (transmission rate and time product) and C represents additional
constraints. Typical constraints for such a reservation are a lower and an upper bound-
ary for the transmission rate. Currently, malleable advance reservations are realized by
computing a deferrable advance reservation which satisfies the requested data size. Fur-
ther processing strategies are under consideration [BMPP07]. Eventually, it is up to the
capabilities of the systems involved which strategies can be used.
4 Integration into multidomain environments
The ARGON system introduced in the previous section allows for an integration of net-
work resources into a Grid environment. To achieve this, ARGON requires detailed knowl-
edge of the underlying network’s topology and must be able to directly control the network.
This is only feasible if the network belongs to a single administrative authority.
Once there are several domains, each controlled by its own NRM, naturally the desire will
arise to provide support for services crossing the domain borders via interdomain links.
Integrating all these domains under a single NRM is rarely a viable solution, since the
possibility to use an existing, customized system will have to remain. Also, an administra-

Domain ADomain A Domain YDomain Y Domain BDomain B User endpoint
Border endpoint
A.B1 Y.B1 Y.B2 B.UB.B2
B.B1A.B2A.U
Figure 2: Illustration of an interdomain reservation scenario
tive authority will most probably insist that only its NRM is allowed to directly configure
the network. In many cases, it will not even want to disclose its internal network topology.
Therefore, an interdomain connection has to be composed by multiple single domain con-
nections that terminate at endpoints connected by interdomain links. Since this introduces
additional complexity, a new system is needed that offers an NRM-like interface for a
metascheduler (cf. Section 3.2) or a human user and that hides the multidomain aspects of
the underlying network.
Such a system is currently being developed in the Phosphorus project funded by the Euro-
pean Union. In the first project stage, the Network Service Plane (NSP) [F+07] that unifies
and abstracts from the underlying domains is implemented as a single, central entity. It is
planned to extend the current solution to allow for a distributed operation in the future.
To integrate several NRMs with different interfaces in the NSP, the concept of an adapter
is used. In general, the NSP does not communicate with a NRM directly, but through
a NRM adapter whose task is to “translate” between the unified NSP interface and the
NRM-specific interface. Conceptually, the adapter code does not belong to the NSP, but
is an extension of the NRM that allows it to be part of a multidomain network. Currently,
NRM adapters exist for ARGON, DRAC, and UCLP.
The interface specification of the NSP interface towards a metascheduler or towards a
human user has been extended from the ARGON interface. It is the same as that of the
NRM adapter interface towards the NSP. The reason is that the services offered by the NSP
are the same services as those offered by a NRM.
The topology information managed in the NSP is comprised of domains, endpoints, and
interdomain links (cf. Figure 2). Endpoints are divided into two different categories: Bor-
der endpoints are used for interdomain links and must therefore be known within the NSP,
but are not of further interest for a metascheduler or for a human user. User endpoints
in contrast are actually used by applications. Therefore, they need not be known within
the NSP, they must merely be locatable, i.e. there must be a mechanism to look up which
domain a user endpoint is located in. In the current NSP implementation, this is achieved
by assigning separate user endpoint address spaces to the domains.
Topology information currently enters the NSP through an administrative interface, the
topology interface. This interface is accessed by the NRM adapters running inside the
domains. They are in charge of keeping the domain-specific information up to date. Cur-
rently, interdomain links are entered manually through an interactive management client.
A protocol to exchange topology information between multiple NSP instances is under

development.
An example of an interdomain reservation is now illustrated for the scenario depicted
in Figure 2. When the NSP receives a request for user endpoints A.U and B.U to be
connected, it is not aware of the resource usage within the domains. It therefore must split
the requested interdomain connection to intradomain connections, each of which fulfils
the constraints specified in the original request. The path computation module would
first return the path A.U–A.B2, B.B1–B.U. The NSP queries the availability of each of
these connections at the corresponding domain’s NRM adapter. In case all connections
are available, the corresponding reservations are made and a positive reply is sent back
to the requester. If, however, resources turn out to be unavailable (e.g. the intradomain
path A.U–A.B2), they are marked accordingly and a new path is calculated, in this case
A.U–A.B1, Y.B1–Y.B2, B.B2–B.U. If no available path can be found, a negative reply is
returned.
5 Conclusions and further work
The ARGON system presented in this paper enables the integration of metro and wide
area network resources into a Grid environment. This is achieved by offering guaranteed
network services via advance reservations. Different types of network connectivity are
supported according to the needs of the applications. The corresponding interface towards
the Grid allows applications or metaschedulers to efficiently plan complex workflows that
require coallocation of heterogeneous Grid resources comprising the network.
ARGON has been designed, developed, and successfully demonstrated in the scope of the
VIOLA and Phosphorus project. In this context ARGON acts as a single-domain NRM
and interfaces to MPLS and GMPLS equipment from different vendors. Furthermore,
ARGON has been successfully integrated into the Phosphorus testbed containing multiple
NRMs which originate from various research projects. The newly developed NSP (cf.
Section 4) connects different NRMs and provides a heterogeneous multidomain solution.
In both cases, the practical experience gathered was and is used to enhance ARGON.
The future work on ARGON and the NSP comprises technical aspects, enhancements of
the Grid services and the corresponding interface, as well as associated theoretical issues.
Furthermore, emerging standards like WS-Agreement and WS-Notification are under con-
sideration, e.g. WS-Notification can be used to distribute information on changing con-
ditions and especially network failure situations in a standardized way. In these cases,
an additional negotiation phase can be triggered in order to modify, postpone, or cancel
reservations.
Finally, also distributed control plane solutions are evolving parallel to the centralized Net-
work Resource Managers. In the Phosphorus project, not only the Network Service Plane
described in Section 4 is being developed, but also an extended GMPLS control plane that
(among other Grid features) is capable of making advance reservations. Integrating such a
distributed approach with the centralized approaches of Network Resource Managers and
the Network Service Plane is a challenging future goal of Phosphorus.

References
[B+07a] C. Barz et al. Co-Allocating Compute and Network Resources - Bandwidth on Demand
in the VIOLA Testbed. In Proceedings of the CoreGRID Symposium, CoreGRID Series.
Springer, September 2007.
[B+07b] Lina Battestilli et al. EnLIGHTened Computing: An Architecture for Co-scheduling and
Co-allocating Network, Compute, and other Grid Resources for High-End Applications.
In Proceedings of 4th International Symposium on High Capacity Optical Networks and
Enabling Technologies (HONET), 2007.
[BBMP08] C. Barz, U. Bornhauser, P. Martini, and M. Pilz. Timeslot based resource management in
grid environments. In Proceedings of the IASTED International Conference on Parallel
and Distributed Computing and Systems (PDCN), February 2008.
[BMPP07] C. Barz, P. Martini, M. Pilz, and F. Purnhagen. Experiments on Network Services for
the Grid. In Proceedings of the 32nd IEEE Conference on Local Computer Networks
(LCN’07), pages 45–54, 2007.
[Bur04] Lars-Olof Burchard. Advance Reservations of Bandwidth in Computer Networks. PhD
thesis, Berlin University of Technology, August 2004.
[C+06] Mauro Campanella et al. Deliverable DJ3.3.2: Functional Specification of GE´ANT2
Inter-domain Manager (IDM) Prototype. Technical report, GE´ANT2, 2006.
[DIC07] DICE (DANTE-Internet2-CANARIE-ESnet) Collaboration. http://www.
geant2.net/server/show/conWebDoc.1308, accessed 31 Dec. 2007.
[F+03] I. Foster et al. The Grid 2: Blueprint for a New Computing Infrastructure-Application
Tuning and Adaptation. Morgan Kaufman, San Francisco, CA, 2003.
[F+07] S. Figuerola et al. PHOSPHORUS: single-step on-demand services across multi-domain
networks for e-science. Proceedings of SPIE, 6784, 2007.
[Fer07] Tiziana Ferrari. Grid Network Services Use Cases from the e-Science Community.
Technical report, Open Grid Forum (OGF), December 2007. OGF Reference: GFD.122.
[FVA06] A. Farrel, J. P. Vasseur, and J. Ash. A Path Computation Element (PCE)-Based Archi-
tecture. IETF RFC 4655 (Proposed Standard), August 2006.
[GO00] Roch Gue´rin and Ariel Orda. Networks with Advance Reservations: The Routing Per-
spective. In Proceedings of the IEEE INFOCOM, pages 118–127, 2000.
[T+05] F. Travostino et al. Project DRAC: Creating an applications-aware network. Nortel
Technical Journal, February 2005.
[T+06a] Atsuko Takefusa et al. G-lambda: coordination of a grid scheduler and lambda path
service over GMPLS. Future Gener. Comput. Syst., 22(8):868–875, 2006.
[T+06b] T.Lehman et al. DRAGON: a framework for service provisioning in heterogeneous grid
networks. IEEE Communications Magazine, 44:84–90, March 2006.
[UCL07] UCLP. http://www.uclp.ca, accessed 31 Dec. 2007.
[VIO07] VIOLA. http://www.viola-testbed.de, accessed 31 Dec. 2007.