Content uploaded by Daniel Mallmann
Author content
All content in this area was uploaded by Daniel Mallmann
Content may be subject to copyright.
2007
German e-Science
Available online at http://www.ges2007.de
This document is under the terms of the
CC-BY-NC-ND Creative Commons Attribution
Requirements and Design of a Collaborative
Online Visualization and Steering Framework for
Grid and e-Science infrastructures
Morris Riedel1, Wolfgang Frings1, Sonja Dominiczak1, Thomas Eickermann1,
Thomas D¨
ussel1, Paul Gibbon1, Daniel Mallmann1, Felix Wolf1, Wolfram Schiffmann2
1John von Neumann Institute for Computing
Central Institute for Applied Mathematics
Forschungszentrum J¨
ulich GmbH, 52425, J¨
ulich, Germany
2Institute of Computer Architecture, Department of Computer Science
University of Hagen, 58097 ,Hagen, Germany
email: {m.riedel}@fz-juelich.de
phone: (+49 2461) 61 3651, fax: (+49 2461) 61 6656
Abstract
Many production e-Science infrastructures (e.g. DEISA, D-Grid) have
begun to offer a wide variety of services for end-users during the past
several years. Many e-Scientists solve their scientific problems by us-
ing parallel computing applications on clusters and collaborative on-
line visualization and steering (COVS) is known as a tool for analyz-
ing and better understanding of these applications. In absence of a
widely accepted COVS framework within Grids, visualizations are often
created using proprietary technologies assuming a dedicated scenario.
This makes it feasible to analyze the usual requirements to provide a
blueprint for a more general COVS framework that can be integrated
into Grid middleware systems such as UNICORE, gLite, or Globus
Toolkits. These requirements lead to a design that was successfully
implemented as a higher-level service in UNICORE and presented at
numerous places such as the Open Grid Forum 19 and 20, Europar
2006, Supercomputing 2006 and DEISA trainings.
1 Introduction
Modern large-scale scientific research often relies on the collaborative use of
a Grid or e-Science infrastructure (e.g. DEISA, EGEE) with computational or
other types of physical resources. One of the major goals of these infrastructures
is to facilitate the routine interaction of scientists and their work with advanced
problem solving tools. Many e-Science applications within these Grids aim at
numerical simulations of physical, chemical, biological processes or other scien-
tific domain-specific problems and parallel computing is widely accepted as an
essential tool for the solution of these problems. In the context of parallel com-
puting, visualization and steering is known as a tool for analyzing and better
understanding of shared parallel applications that run on a cluster or supercom-
puter. The concept to visualize complex scientific datasets (e.g. vectors, arrays)
2 Morris Riedel et al.
lead often to more insights in the computational process of the application. Fur-
thermore, a wider range of control is given through steering of the application by
influencing its parameters during runtime. Non photo-realistic visualizations are
used, since the rather schematic visual representations are often able to convey
more information augmented with specific details such as scales or numbers that
improve the insights of scientists.
At the time of writing, the major Grid middleware systems that provide
access to computational resources offer no interactive services for schematic col-
laborative online visualization and steering (COVS) of parallel simulations, even
if this technique is well established among scientists. This motivates this work
by identifying the necessary functionality and quality of such a higher-level Grid
service. Such service enables scientists to observe the intermediate steps dur-
ing the computation of the simulations (online visualization) and can interact
with the parallel simulation at once to influence its computation (computational
steering). The integration of these services into Grids leads to several benefits
for end-users, including single sign-on (1 times password provisioning) and vir-
tualization of resources (no knowledge of hostname and username details). In
absence of a widely accepted framework for COVS services, the most visualiza-
tions are often created by some proprietary mechanisms.
The lack of a common framework for COVS and standardized methods to
create an online connection between a simulation and visualization is contrary
to fundamental design principles of software engineering. This paper identifies
requirements for a proposed design of a COVS framework within e-Science infras-
tructures. This includes the addressing of challenges in the area of visualization,
communication, collaboration, as well as network and Grid issues. This require-
ment analysis lays the foundation for a general design of a COVS framework and
its reference implementation using UNICORE [3] and VISIT [7].
The remainder of this paper is structured as follows. In Section 2 we define
several requirements for a COVS framework in Grids. Section 3 presents the
general design of the COVS framework, while Section 4 describes its implemen-
tation within UNICORE Grids and use case scenarios. The paper ends with a
survey of related work and some concluding remarks.
2 Requirements for a COVS Framework in Grids
The integration of a COVS framework into e-Science infrastructures implies
that specific requirements of these environments must be addressed. In addi-
tion, a COVS framework must satisfy all specific requirements for a scientific
visualization and steering tool. The addressing of all these requirements lay the
foundation for a COVS framework that allows for scientific visualization, de-
fined by R.B. Haber et al. in [12]. In this sense, scientific visualization is the
use of computer imaging technology as a tool for understanding, analyzing, and
illustrate complex and comprehending data typically obtained by parallel simu-
lations. Lessons learned from e-Science in the context of COVS indicate that the
demands for visualization are increasing with the number of new applications in
Grids and often conflicting requirements for a COVS framework.
GES 2007 Collaborative Online Visualization and Steering Framework 3
2.1 Challenges in Visualization and Communication
A COVS framework should support a wide variety of existing visualization
technologies (e.g. VTK, AVS/Express) in order to circumvent duplicate efforts
in re-developing them and to gain the experience from the visualization com-
munity. The integration of such systems into a COVS framework provide the
functionality to create visualization idioms defined by R.B. Haber et al. in [12].
A visualization idiom is a specific sequence of data enrichment and enhance-
ment transformations, visualization mappings, and rendering transformations
that produce a schematic display of a scientific dataset.
Another requirement is the support of communication libraries for visualiza-
tion and steering (e.g. VISIT [7], PV3 [13]) that are used to transport datasets
between the visualization and simulation and convey additional information such
as steering commands. Hence, these powerful technologies decouple the visual-
ization from the simulation and also decouple both from communication issues.
Another fundamental challenge for a COVS framework is to allow for online
visualization. This implies that the framework must transfer scientific data be-
tween the simulation and visualization in a stepwise fashion to allow for online
visualization of single computational steps and thus for computational steering.
The demand to use parallel computational resources as effective as possible leads
to the requirement of adding computational steering technologies (e.g. VISIT,
gViz [1], ICENI [16]) into the COVS framework. This enable scientists to focus
on special areas of computation or to early abort computations that turn out
to be false during its visualization. Finally, a COVS framework requires a se-
cure bi-directional data transfer between the visualization and simulation with
high performance, because scientific data computed by the simulation flows from
the simulation to the visualization, while the steering data defined within the
visualization flows from the visualization to the simulation.
In this context it seems reasonable to take the underlying network infras-
tructure into account that should provide an adequate bandwidth to support
fast connections that in turn realize real-time behaviour of online visualizations.
DEISA and TeraGrid, for instance, provide perfect bandwidth capabilities due
to dedicated connections. Also, to use steering as effective as possible, a COVS
framework should provide protocols with minimal overhead in order to use bi-
directional connections with low latencies during COVS sessions.
2.2 Requirements for Collaborative Sessions
This section emphasize on the collaborative nature of COVS sessions with ge-
ographically dispersed participants that leads to the distinction of separate roles.
First and foremost, a person that use the COVS framework is in the participant
role if the person shares the view on one visualization of a parallel simulation
with all other n-1 participants. While some people only act in the participant
role, there are other people that may represent more than one role. This im-
plies that the functionality of the COVS framework for one role differs from the
functionality offered to other roles. For instance, a person that use the COVS
framework is the master of the COVS session if this person uses the framework
4 Morris Riedel et al.
to submit and control a parallel simulation that runs on a computational Grid
resource. Hence, other participants do not need to submit a simulation job.
A person in the approver role uses the COVS framework and makes decisions
which participants are allowed to join COVS session. Thus, a person that uses
the COVS framework to apply for participation in a COVS session is named a
candidate and a person in this role needs approval by a person in the approver
role to become a participant of a COVS session. From this it follows that some
candidates may not get an approval since they are not allowed to share the view
in the COVS session. Also, the master role is able to explicitly exclude partici-
pants from the session, if their behavior or technical reasons require such rather
aggressive actions. In both cases the person acts in the blocked participant role.
Furthermore, capabilities to steer a parallel simulation during a collaborative
session raises the requirement for a mechanism of mutual exclusion of partici-
pants during steering. Hence, only one participant in the steerer role is allowed
at the same time to steer a parallel simulation during a COVS session in order
to ensure the consistency of the simulation and its computation. In addition,
only participants in the collaborator role are allowed to change the view of the
visualization, which is not necessarily steering of the simulation itself, e.g. in
case of turning the viewpoint of already computed data by 45 degrees.
Also, other challenges include the distribution of data and control informa-
tion. A COVS framework requires a multiplexer entity that multiplexes the
output of 1 parallel simulation to n bi-directional connections that provide n
online visualizations of participants with the same data. This in turn raise a
demand for scalable multiplexing in the sense that n participants can still share
the same view of the data without an appreciable loss in quality of the scientific
visualization. Second, a COVS framework requires an collaboration entity that
transports the collaboration data (e.g. change level of detail, colors) from 1 visu-
alization to all the other n-1 visualizations to ensure that all participants in the
COVS session share the same view on the data. Thus, if one visualization makes
a turn of 180 degrees, all other visualizations should also turn 180 degrees.
Finally, several requirements are related to the scope of session control. The
dynamics of collaborations among the scientists lead to the demand of dynam-
ically attaching and detaching participants to the visualization during the run-
time of the simulation without influencing the quality of visualizations from other
participants. This in turn raises a demand to monitor the status of participants.
Thus, in order to determine whether all participants of a COVS session are
already connected or still connected to the parallel simulation, the COVS frame-
work requires a monitor mechanism that provides status information about the
connection of all participants. This monitor mechanism should also include per-
formance statistics of the bi-directional connections to all n visualizations of the
participants in order to detect those that represent bottlenecks of a collaborative
session and thus influence the overall quality of the session.
Finally, the overall management of a COVS session requires authorized ses-
sion management control actions within the COVS framework that include the
addition or removal of participants that may represent bottlenecks.
GES 2007 Collaborative Online Visualization and Steering Framework 5
3 Design and Implementation of the COVS Framework
The design presented here provides an architectural blueprint of a COVS
framework that can be implemented in different Grid environments that typi-
cally are based on different Grid middleware systems (e.g. UNICORE, gLite, or
Globus Toolkits). One of the key considerations of this framework is to define
an architecture that works for a wide variety of applications in the scientific
domain. The design requirements identified in the previous section are most
crucial to the framework design. A COVS framework that addresses them is
more likey to achieve high levels of design, component and thus code reuse by
still providing much flexibility. The general design and the implementation pre-
sented here satisfy all the described requirements by using UNICORE as a Grid
middleware and VISIT as the communication library. This section highlights
several crucial framework parts in more detail while an overview of the COVS
framework design is illustrated in Figure 1.
Figure 1: COVS architecture for UNICORE, gLite or Globus Toolkits as Grid
middleware. Web services are used for communication as well as SSH tunnels
for the bi-directional transfer of scientific data and steering commands.
6 Morris Riedel et al.
An adequate framework for COVS that is built on Grid computing and com-
munication technologies relies on technologies provides by the Grid and e-Science
infrastructure. In this context, the COVS framework is based on Grid middle-
ware systems to gain all the benefits of these systems and thus of the whole
infrastructure. Hence, the COVS framework is integrated seamlessly into the
Grid or e-Science infrastructure by the meaning of hiding the fact that resources
are physically distributed and tpically managed via Resource Management Sys-
tems (RMS) such as Torque, LSF or LoadLeveler. Such a transparency includes
differences in security policies, data representation and how a resource is accessed
when using the Grid services of the COVS framework.
The Grid middleware represents a crucial core building block for simulation
job management as shown in Figure 1. It is capable of submitting, controlling
and managing computational jobs such as a parallel simulation that run on su-
percomputers or clusters. This is mainly supported by core Grid services such as
the UNICORE Atomic Services (UAS) [14] or OGSA - Basic Execution Services
(BES) [4]. However, the Grid middleware must also be extensible to allow for the
creation of additional Grid services that represent higher-level services for COVS
sessions. In particular, the design of the framework relies on the Web Services
Resource Framework (WS-RF) [9] standard for the development of such services.
In more detail, the WS-RF compliant COVS Grid service represents another core
building block and is used for COVS session management. The COVS Factory
service implements the WS-RF factory pattern [9] and brings COVS Session
Resources with properties into existence that are in turn accessed by the COVS
Session service. As shown in Figure 1, the COVS Session service interacts with
the multiplexer and collaboration server and exposes the status of the COVS
session as WS-Resource properties [9]. This allows the COVS Grid service to
have the explicit control over the scientific dataflow to numerous participants
and the collaboration data exchange. To sum up, the COVS framework relies
on middleware that offers services via open connection technologies according to
standard rules or emerging Grid standards such as WS-RF. We implemented the
COVS Grid service within the WS-RF based UNICORE 6 Grid system [11].
The integration of COVS functionality into Grids implies that the funda-
mental authorization and authentication of end-users is managed by the corre-
sponding Grid middleware. In the context of UNICORE, only end-users that
are configured within the UNICORE user database [3] are able to participate in
a COVS session. Of course, the seamless integration into the security infrastruc-
ture in Grids places the requirement on the COVS framework to preserve single
sign-on. In more detail, the design of the COVS framework relies on SSH for
the bi-directional data transfer and thus a typical end-user must know a con-
crete hostname, username of the remote host to establish the connection. But
to preserve the single sign-on the Grid middleware provides all these necessary
details and thus retains transparency to end-users. As shown in Figure 1, the
Grid middleware client and the scientific visualization interact with each other.
This interaction is used to transfer all necessary details to establish an SSH con-
nection from the scientific visualization to the parallel simulation. In the context
GES 2007 Collaborative Online Visualization and Steering Framework 7
of the implementation with UNICORE, we use an RSA-based authentication. A
SSH session key is generated in the client and transferred to the target system
via UNICORE. Next, the public key is inserted into the authorizedkeys file of
the target system and thus the client tier gets access per SSH. When the visu-
alization session is finished, the key is removed. This approach is similar to the
SSH interactive access to UNICORE as described by Riedel et al. in [5].
E-scientists use Grid resources with parallel computing techniques to solve
problems in their areas of science. Today, most parallel computers of supercom-
puting centers are totally booked out or their demand is five times higher than
what the resources can offer. Thus, computational time on parallel supercom-
puters is not cheap and should be used as effectively as possible. In this context,
the COVS framework must support a minimization of the load on a steered
parallel simulation and the circumvention of failures or slow operations by the
visualization that may disturb the simulation progress. We achieved that by us-
ing the VISIT communication library. In VISIT, the simulation acts as a client
which initiates all operations like opening a connection through the established
SSH tunnel, sending data to be visualized or receiving new steering parameters.
Also, we identified a demand for a managing Grid client (e.g. GPE Client
suite [8], or CogKits ) within the COVS framework that provides an overview of
the session by using information exposed by the underlying Grid middleware.
Figure 2: COVS GridBean for collaborative session management. End-users can
use popup menus to conveniently monitor and manage COVS sessions.
8 Morris Riedel et al.
Therefore, the Grid client as well as its COVS specific plugin for COVS
session management also represent core building blocks. In our implementation,
we used the GPE application client and developed a COVS GridBean as shown
in Figure 2. GridBeans are scientific application-specific plugins for the GPE
clients [8]. Finally, the parallel simulation and scientific visualization represent
core building blocks that are scientific application-specific and thus it is feasible
provide an concrete example in the next Section.
4 Use Case Scenarios
In order to demonstrate the re-usability for parallel simulations and scientific
visualizations, we provide a concrete use case scenario that is based on the sci-
entific visualization Xnbody [15]. It is based on VTK and integrates the VISIT
toolkit as shown in Figure 3. Xnbody shows the output of parallel simulations
in the context of n-body problems by using the COVS framework implemen-
tation in UNICORE. In particular Xnbody is used in the context of plasma
physics (PEPC simulation code) and astrophysics (nbody6++ code). Without
using the COVS framework implementation end-users have to manually pro-
vide details about the SSH connections to remote sites. When using the COVS
framework implementation UNICORE provides all necessary details for the SSH
connection and thus provides transparency and single sign-on to end-users of
Xnbody, including the session management capabilities of the COVS GridBean.
Figure 3: Xnbody uses the ’use UNICORE’ checkbox to seamlessly connect to
supercomputing sites via security information based on UNICORE.
GES 2007 Collaborative Online Visualization and Steering Framework 9
5 Related Work
There is a wide variety of related work in the field of visualization and steer-
ing within Grids. For instance, Brodlie et al. describe in [10] a framework for
distributed and collaborative visualization and how it can be potentially imple-
mented by several visualization systems. But the framework is rather high-level
and thus we followed an approach that was closer to current production Grids.
Furthermore, there is considerable research interest in new visualization and
steering technologies within many national and international Grid initiatives.
First and foremost, Kleijer et al. describe in [2] the API for the Grid-based
visualization systems of the NAREGI Grid infrastructure. This API consists of
a visualization library and a Grid visualization service API that provides Grid
service functionality conform to the Open Grid Services Architecture (OGSA).
In the last years, the API evolved to a wide-variety of WS-RF compliant services
for visualization such as a post-processing service. Even if WS-RF is also used
in our approach, we use a different approach since the NAREGI services are
rather image-based instead of following the online visualization technique. That
means that the scientific data as well as their rendering and visualization are
completely computed within the Grid that finally leads to a compressed image
that is transferred to clients and thus makes it difficult to apply steering.
K¨
ockerbauer et al. describe in [6] the visualization system that is used in
the context of the Austrian Grid. This system is named as Grid Enabled Vi-
sualization Pipeline (GVID) and provides high quality Grid-based visualization
of scientific datasets on thin clients (e.g. playstation clients). In more detail,
the scientific data is efficiently encoded with the H262 code into a video stream
and transferred to the client afterwards. The client in turn decodes the video
stream for visualization and can apply steering commands that are tracked by
the GVID Event-Encoder. Finally, GVID is used for Galaxy visualizations, but
the major difference to our approach is that it is not seamlessly integrated as a
higher-level service into a specific Grid middleware.
The UK RealityGrid project focused on how scientists can make more ef-
fective use of a Grid and its visualization resources. The most known work in
RealityGrid is around its steering library that enable calls which can be embed-
ded into each of the three components of its architecture, namely simulation,
visualization, and a dedicated steering client. More recently, older prototypes of
RealityGrid are renewed towards OGSA environments and unfortunately tightly
integrated into the Imperial College e-Science Networked Infrastructure (ICENI)
[16]. Inversely, the COVS framework is rather loosely coupled from the under-
lying Grid middleware.
6 Conclusions
The design of the COVS framework provides a sophisticating blueprint for
implementations in different Grid middleware systems. Thus, to demonstrate
that the COVS requirements and problems addressed within this paper are of
10 Morris Riedel et al.
practical relevance, we implemented a COVS framework implementation in UNI-
CORE by using the VISIT toolkit. This implementation is ready for production
and can be used by all VISIT-enabled simulations and visualizations. The seam-
less integration of the COVS framework into UNICORE Grids improves the
work of scientists by providing online schematic visualizations of complex paral-
lel simulations by still retaining the single sign-on feature of Grid and e-Science
infrastructures and secure data transfers with SSH across Grids.
Finally, the work within this paper was successfully demonstrated at the
OGF18, Europar 2006, Supercomputing 2006, and recently at the visualization
workshop at OGF19. Furthermore, it was demonstrated to end-users in DEISA
at the DEISA training in November 2006 and is continously shown as one ex-
ample of an higher-level service in UNICORE at various UNICORE presenta-
tions world-wide. Nevertheless, deploying the proposed COVS architecture is an
important next step to broadly incorporate implementations of a COVS frame-
work into production Grid environments. Once an implementation of the COVS
framework is deployed within production Grids such as DEISA or D-Grid, an
important tool for an efficient use of the Grid is accomplished.
References
1. K.Brodlie et al. Visualization in Grid computing environments: Technologies,
Applications and Challenges. In Computer Graphics Forum, Vol.24(2), 2005
2. P. Kleijer et al. API for Grid Based Visualization Systems. In GGF 12 Workshop
on Grid Application Programming Interfaces, 2004
3. A. Streit et al. UNICORE - From Project Results to Production Grids, Elsevier,
Grid Comp. and New Frontiers of High Perf. Proc., pages 357–376, 2005
4. OGSA - BES, http://www.ogf.org/gf/group info/view ˙php?group=ogsa-bes-wg.
5. M. Riedel et al. Enhancing Scientific Workflows with Secure Shell Functionality
in UNICORE Grids, In Int. e-Science conference, Melbourne, 2005
6. T. K¨
ockerbauer et al. GVid - Video Coding and Encryption for Advanced Grid
Visualization. In Proc. of the 1st Austrian Grid Symposium, Linz, 2005
7. Visualization Interface Toolkit, http://www.fz-juelich.de/zam/visit
8. R. Ratering et al. GridBeans: Supporting e-Science and Grid Applications. In
2nd Int. e-Science conference, Amsterdam, 2006,
9. OASIS - WS-RF Technical Committee,
http://www.oasis-open.org/committees/tc home.php?/wg abbrev=wsrf
10. K.Brodlie et al. Distributed and Collaborative Visualization. In Computer Graph-
ics Forum, Vol.23, 2004
11. UNICORE Grid Middleware, http://www.unicore.eu
12. R.B.Haber and D.A.McNabb. Visualization idioms: A conceptual model for sci-
entific visualization systems, In Vis. in Scientific Comp., pages 74-93
13. PV3, http://www.raphael.mit.edu/pv3
14. M. Riedel et al. Standardization Processes of the UNICORE Grid System. In
Proc. of 1st Austrian Grid Symposium 2005, Linz, Austria, pages 191-203
15. Xnbody - Visualization, http://www.fz-juelich.de/zam/xnbody
16. J. Cohen et al. RealityGrid: An Integrated Approach to middleware through
ICENI. In Phil. Transactions of the Royal Society, 363, pages 1817-1827, 2005
