|
Scheduling for Numerical Linear Algebra Library at Scale
Jack
Dongarra
Innovative
Computing Laboratory
Computer
Science Dept.
University of Tennessee
Knoxville, TN, U.S.A.
In this
talk we will look at some of the issues numerical library developers are
facing when using manycore systems with millions of threads of execution.
Back to Session I
|
|
Clouds and ManyCore: The Revolution
Daniel A. Reed
Microsoft Research
Redmond, Seattle, U.S.A.
As Yogi
Berra famously noted, “It’s hard to make predictions, especially about the
future.” Without doubt, though, scientific discovery, business practice and
social interactions are moving rapidly from a world of homogeneous and local
systems to a world of distributed software, virtual organizations and cloud
computing infrastructure. In science,
a tsunami of new experimental and computational data and a suite of increasingly
ubiquitous sensors pose vexing problems in data analysis, transport,
visualization and collaboration. In society and business, software as a
service and cloud computing are empowering distributed groups.
Let’s
step back and think about the longer term future. Where is the technology
going and what are the research implications?
What architectures are appropriate for 100-way or 1000-way multicore
designs? How do we build scalable
infrastructure? How do we develop and support software? What is the ecosystem of components in
which they will operate? How do we optimize performance, power and
reliability? Do we have ideas and
vision or are we constrained by ecosystem economics and research funding
parsimony?
Biographical Sketch
Daniel A. Reed is Microsoft’s Scalable and Multicore Computing
Strategist, responsible for re-envisioning the data center of the
future. Previously, he was the
Chancellor’s Eminent Professor at UNC Chapel Hill, as well as the Director of
the Renaissance Computing Institute (RENCI) and the Chancellor’s Senior
Advisor for Strategy and Innovation for UNC Chapel Hill. Dr. Reed is a member of President Bush’s
Council of Advisors on Science and Technology (PCAST) and a former member of
the President’s Information Technology Advisory Committee (PITAC). He recently chaired a review of the federal
networking and IT research portfolio, and he is chair of the board of
directors of the Computing Research Association.
He was previously Head of the Department of Computer Science at the University of Illinois at Urbana-Champaign
(UIUC). He has also been Director of the National Center for Supercomputing
Applications (NCSA) at UIUC, where he also led National Computational Science
Alliance. He was also one of the principal investigators and chief architect
for the NSF TeraGrid. He received his
PhD in computer science in 1983 from Purdue University.
Back to Session I
|
|
Present and future challenges as
we architect for the Exascale
Alan Gara
Dept. of Computer Science
Indiana University
U.S.A.
In this
presentation current trends toward achieving Petascale computing are
examined. These current trends will be contrasted with what is needed to
reach the Exascale. Possible directions and critical enabling technologies
will be discussed.
Back to Session I
|
|
Effective computing on heterogeneous platforms
Anne
Trefethen
Oxford University
Oxford, U.K.
We have entered an era where at every scale of
computing - desktop, high-performance and distributed - we need to deal with
heterogeneity. Systems are made up of
multicore chips and accelerators in an assortment of hardware architectures
and software environments. This has
created a complexity for scientific application developers and algorithm
developers alike. Our focus is on
effective algorithms and environments across these scales to support
efficient scientific application development.
Back to Session I
|
|
The Evolution of Research and Education Networks
and their Essential Role in Modern Science
William E. Johnston
Senior Scientist and Energy Sciences Network (ESnet) Department Head
Lawrence Berkeley National Laboratory
Berkley, CA, U.S.A.
In the
past 15 years there has been a remarkable increase in the volume of data that
must be analyzed in world-wide collaborations in order to accomplish the most
advanced science and a corresponding increase in network bandwidth,
deployment, and capabilities to meet these needs. Further, these changes have
touched all aspects of science including, in addition to data analysis,
remote conduct of experiments and multi-component distributed computational
simulation.
Terabytes
of data from unique and very expensive instruments must be collaboratively
analyzed by the many science groups involved in the experiments. The highly
complex, long-running simulations needed to accurately represent macro-scale
phenomenon such as the climate, stellar formation, in-vivo cellular functioning
in complex organisms, etc., all involve building applications that
incorporate and use components that are located at the home institutions of
many different scientific groups.
The
volume of traffic in research and education networks has increased
exponentially since about 1990. Virtually all of this increase – demonstrably
so in the past five years – is due to increased use of the network for moving
vast quantities of data among scientific instruments and widely distributed
analysis systems, and among supercomputers and remote analysis centers.
Further, this data movement is no longer optional for science: Increasingly
large-scale science is dependent on network-based data movement in order for
the science to be successful.
Modern
science approaches require that networks provide not only high bandwidth, but
also advanced services. Scheduled and on-demand bandwidth enables connection
and simultaneous operation of instruments, local compute clusters,
supercomputers, and large storage systems. Low latency, high bandwidth,
secure circuits interconnect components of simulations running on systems
scattered around the country and internationally. Comprehensive, global
monitoring and reporting that allow distributed workflow systems to know
exactly how end-to-end paths that transit many different networks are
performing. At the same time, the network must provide a level of reliability
that is commensurate with the billion dollar instrument systems, scarce
supercomputers, and the hundreds of collaborating scientific groups being
interconnected that is typical of large-scale science.
In this
talk I will look at how network architectures, technologies, and services
have evolved over the past 15 years to meet the needs of science that now
uses sophisticated distributed systems as an integral part of the process of
doing science. One result of this is that the R&E community has some
unique communications requirements and some of the most capable networks in
the world to satisfy those requirements. I will also look at the projected
requirements for science over the next 5 to 10 years and how the R&E
networks must further expand and evolve to meet these future requirements.
Back to Session I
|
|
Grids, Clouds and HPC: Opportunities and Challenges
Dr.
Frank Baetke - Global HPC Technology Program Manager
Richardson, TX, U.S.A.
New trends in the HPC area can be derived
from increasing growth-rates at the lower end of the market, specifically at
the workgroup and departmental level, and from concepts which are based on
the original promises of computational grids. Those trends combined with the
ever increasing demand for even higher
component densities and higher energy efficiency generate additional
challenges: examples of new products will be shown which specifically address
those issues.
Back to Session II
|
|
Intel - Delivering Leadership HPC
Technology Today and Tomorrow
Stephan Gillich
Director HPC EMEA
Enterprise Marketing EMEA
Intel GmbH, U.S.A.
We are
excited about the opportunity that lies in front of us as our
manufacturing processes move from 45nm to 32nm to 22nm and on to 16nm.
This progress will e.g enables us to pack more cores into each
processor. But it is not just about adding cores that will deliver the
performance, it's about creating the right infrastructure to keep all
those cores busy which is just as valuable as the processor and the
infrastructure - combining the energy efficiency and performance gains
in hardware with advances in software and programming techniques. It is
about providing the SW tools that enable organizations to create
sustainable high performing applications that scale in performance.
This talk will cover Intel's HPC technology fueled by our manufacturing
leadership and driven by our research - providing a robust and powerful
infrastructure. It will conclude with an outlook into Intel's Tera-scale
research program.
Back
to Session II
|
|
High Performance
Storage Solutions from DataDirect Networks
Toine Beckers
DataDirect Networks
Inc., Netherlands
With the growing needs for High Performance
Computing clusters (from GFlops to TFlops and even PFlops systems) in many
application fields also the need for more and more data storage capacity
increases as well. This often leads to complex, difficult to manage storage
solutions. With the Silicon Storage Appliance products from DataDirect
Networks an easy to manage, scalable and high performance solution is
provided which is becoming widely accepted in the High Performance Computing
Community.
Back
to Session II
|
|
Next-Generation Cluster Management with ClusterVisionOS
Martijn
De Vries
ClusterVision BV
Amsterdam, Netherlands
Setting up and managing a large cluster can
be a challenging task without
the right tools at hand. With the upcoming release of ClusterVisionOS 4,
ClusterVision is introducing the next-generation of its widely used
cluster operating system. The new release of ClusterVisionOS features a
brand new SOAP-based cluster management infrastructure that allows
administrators to monitor and control all aspects of their clusters
through graphical and command-line based interfaces.
In this presentation, various aspects of the
ClusterVisionOS cluster
management infrastructure and associated tools will be described.
Back to Session II
|
|
Green Scalable High Performance Supercomputing
Fabrizio
Magugliani
EMEA
Business Development Director Sicortex
Maynard, MA, U.S.A.
As CPU speeds have reached a point where
simply increasing the clock
frequency is no longer an option, the world has turned to multi-core
CPUs in an attempt to continue to get more computational cycles out
essentially the same system arhitecture. This has led to an increasing
inbalance
between brute computational capability and other aspects of the system.
By coupling an energy efficient CPU core design with an very fast and
reliable interconnect, all on a single piece of silicon, we have
created a low-power/low-heat system that balances computational capability with
high performance I/O. This system provably scales to thousands of
processors
in about one cubic meter of space at 18 KW.
Back to Session II
|
|
The Godson-3 multi-core CPU and its application in High Performance Computers
Weiwu
Hu, Xianggao, Yunji Chen
Institute
of Computing Technology, Chinese Academy of Sciences
Beijing, China
Godson-3
is a multi-core processor based on the 64-bit superscalar Godson-2 CPU core.
It takes a scalable CMP architecture in which processors and global addressed
L2 cache modules are connected in a distributed way and coherence of multiple
L1 copies of the same L2 block is maintained with a directory-based cache
coherence protocol.
The Godson-2
CPU core is a four-issue, out-of-order execution CPU which runs the MIPS64
instruction set. The latest Godson-2F which integrates the Godson-2 CPU
core, 512KB L2 cache, 333MHz DDR2 controller and PCI/PCIX controller achieves
1GHz based on the 90nm STMicro CMOS technology, and has been volume produced
and shipped to the market for low-cost PCs, low-end servers, and many
embedded applications. The CPU core of Godson-3 is enhanced from Godson-2 CPU
core to support efficient X86 to MIPS binary translation, and to optimize
performance, power consumption, reliability and debug methods.
Godson-3
adopts two-dimension mesh topology. Each node in the mesh include an 8*8
crossbar which connects four processor cores, four shared L2-cache banks and
four adjacent nodes in the East, South, West and North. A 2*2 mesh network
can connect a 16-core processor, and a 4*4 mesh network can connect a 64-core
processor. The distributed on-chip L2 cache modules are globally addressed.
Each cache block of L1 cache has a fixed L2 cache home node in which the
cache directory is maintained by directory-based cache coherence protocol.
Each node has one (or more) DDR2 memory controller. IO controllers are
connected through free crossbar ports of boundary nodes.
Based on
the Godson-3 architecture, several product chips are defined and will be
physically implemented. The 4-core Godson-3 chip is designed and fabricated
based on 65nm STMicro CMOS technology. It includes one 4-core node, 4MB L2
cache, two DDR2/3 ports, two HT1.0 ports, two PCIE ports, one PCI port and
one LPC port. It will be taped out in first half of 2008.
One
important application of Godson-3 is the low cost high performance computers
(HPC). Based on Godson-3, the design of one national PetaFLOPS HPC and one
personal TeraFLOPs HPC are planed. This presentation will introduces the HPC
plans based on the Godson-3 multi-core processor.
Back to Session III
|
|
Aggressively
Threaded Systems: A Wise Choice for HPC
Rick Hetherington
Vice President, CTO
Microelectronics
Sun Microsystems, Inc
U.S.A.
Niagara technology, in its infancy,
targeted the commercial computing market.
These
throughput workloads were not very computationally intensive but demanded
memory subsystems that provided high bandwidth and high capacity.
The
second and third generations of NIagara added greatly increased computation
capability to the processing cores while continuing to focus on high
throughput.
The result
is a set of products that efficiently deliver high levels of computational
throughput.
This talk
will discuss the UltraSparc T2 and T2+ processor designs as well as an
analysis of their behavior while executing 'technical' workloads.
Back to Session III
|
|
Managing resources dynamically in SVP – from many-core to Grid
Chris
Jesshope
Professor
of Computer Systems Architecture
University of Amsterdam
Amsterdam, Netherlands
Our computer
systems are becoming pervasive and ubiquitous. It is now
possible for everyday consumer products to have networks comprising
thousands of nodes and Grids will multiply this by another large
factor. These systems may comprise a huge number of heterogeneous
computing elements and will certainly evolve. How can such complex
systems be programmed, when we may not know which nodes are available
at any time. Taking advantage of the huge computing power offered by
this collaboration of elements will require the dynamic management of
concurrency and this is a significant challenge. To solve these
issues, a disruptive approach is promoted in the AETHER European
project, which embeds self-adaptivity at each level of the system,
giving autonomy to the components and enabling the application designer
to concentrate on the application instead of having to cope with all
possible events in the lifetime of a computing resource in such a
rapidly evolving environment. For this purpose, we have introduced the
SANE concept (Self-Adaptive Networked Entity) and a
programming model based on a SANE Virtual Processor (SVP). This provides
a dynamic model of concurrency based on dynamically binding resources
to families of blocking threads in a hierarchical manner. This approach
presents a new dynamic architecture and protocols to enable the sharing
of resources and the consequent management of concurrency. This
presentation will describe the resource management protocols that
enables delegation of work. SANEs are autonomous and from time to time
may be given jobs to execute; a local user may
submit a job or one may be delegated from its environment. In the
latter case, the SANE will have contracted with an external thread to
run that job and to meet certain expectations in its execution, for
example performance. The contract is negotiated using a credit
exchange, where the cost of executing a job is initially assumed to be
the energy expended by the contracted SANE but which can be miodified
by market forces. Thus the contracting thread, which may be acting on
behalf of another SANE, transfers credit for the agreed amount of
energy to execute the work on the contracted SANE. In response, the
contracted SANE agrees to meet the deadlines or performance constraints
imposed by the contracting SANE.
Back to Session III
|
|
Nondeterministic Coordination
using S-Net
Prof
Alex Shafarenko
Department
of Computer Science
University of Hertfordshire
Hatfield, AL10
9AB, UK
Coordination languages have been used for many years
in order to separate computation and concurrency/communication, that is
coordination, concerns. Despite that, a typical coordination language
intrudes into the computational part of the code even though it provides
some abstract projection of those distributed computing realities. As a
result, units of an application program become barely readable in isolation,
without having the "big picture" in mind --- and that big picture
in turn is overburdened with interface details.
We believe that the reason why coordination has
these problems is that true separation between computation and concurrency
concerns is only possible using a nondeterministic glue. Indeed deterministic
coordination abstracts application code as a state-transition system,
introducing synchonization over and above the inimum needed for correct
functioning of the application code. Nondeterministic coordination,
which we describe in this paper, leans towards loose, data-flow-style
composition using asynchronous computational structures ---
and synchronisers where necessary to ensure that the correct data
sets are worked on by fully encapsulated application code units.
The paper will present a coordination language
S-Net, developed and implemented by the authors.
The language is very compact, only using 4
combinators acting on user-defined boxes to create hierarchical networks of
asynchronously communicating components. The boxes are written in a
conventional language and use a conventional stream interface for output,
while the input comes as a standard parameter list.
We expect ordinary engineers to be able to provide
these components. There is only one special box which the user cannot
create and which comes with the S-Net language: the synchrocell. The
significant expressive power of coordination in such a small language
is achieved by using a sophisticated type system with
subtyping, which influences the network "wiring" provided by
the combinators. The coordination program is thus a large algebraic formula
using the combinators, or several such formulae, and it is written by a
concurrency engineer who needs
no detailed knowledge of the application
domain.
Concurrency and self-adaptivity of S-Net is
helped by the fact that user-defined
boxes are assumed to be without
persistent state, i.e. after the output stream has been flushed
and the box terminates, all local state is destroyed, so that the
next invocation of the box can take place at a different location in
the distributed system. Synchrocells retain their state between
invocations but they do not perform computations and consequently
consume no computing power.
In conclusion, we will briefly dwell on the recent
success in applying S-Net to a signal processing problem in radar systems
industry at Thales Research & Technology, France.
Back to Session III
|
|
Fault Tolerance for PetaScale
Systems: Current Knowledge, Challenges and Opportunities
Franck Cappello
INRIA
France
The
emergence of PetaScale systems reinvigorates the community interest about how
to manage failures in such systems and ensure that large applications
successfully complete. Existing results for several key mechanisms associated
with fault tolerance in HPC platforms will be presented during this talk.
Most of
these key mechanisms come from the distributed system theory.
Over the
last decade, they have received a lot of attention from the community and
there is probably little to gain by trying to optimize them again. We will describe
some of the latest findings in this domain.
Unfortunately,
despite their high degree of optimization, existing approaches do not fit
well with the challenging evolutions of large scale systems. There is room
and even a need for new approaches. Opportunities may come from different
origins like adding hardware dedicated to fault tolerance or relaxing some of
the constraints inherited from the pure distributed system theory. We will
sketch some of these opportunities and their associated limitations.
Back to Session III
|
|
Ultra Low Power HPC --- scaling
supercomputing by three orders of
magnitude
Satoshi
Matsuoka
Tokyo
Institute of Technology
Tokyo, Japan
Low power
supercomputing as represented by various power efficient architectures such
as IBM BlueGene and power aware methods are starting to receive considerable
attention in the light of global agenda to reduce energy consumption and also
to alleviate increasing heat density problems. Our new project, Ultra Low-Power
HPC, greatly extend this horizon by taking the innovative approaches to
fundamentally slash energy consumption of supercomputing by up to 3 orders of
magnitude in 10 years. This is achieved by the comprehensive use of new
energy-efficient hardware devices and power-saving algorithms that are
modeled and optimized in a systemwide fashion. Early results from the project
are exhibiting good results in achieving 10-100 times energy efficiency,
mostly by the use of acceleration and new memory device technologies.
Back to Session III
|
|
HPC Interconnection Networks – The
Key to Exascale Computing
Jeffrey Vetter
Oak Ridge National Laboratory and Georgia
Institute of Technology
Interconnection
networks play a critical role in the design of next generation HPC
architectures and the performance of important applications. Despite the
significance of interconnects, current trends in HPC interconnects do not
appear to fulfill the requirements for next generation multi-petaflop and
exaflop systems. Application requirements drive networks with high bandwidth,
low latency, and high message rate, while practical constraints, such as
signaling, packaging, and cost, limit improvements in hardware bandwidth and
latencies. To address these
challenges, Sandia and Oak Ridge National Laboratories have established the
Institute for Advanced Architectures and Algorithms (IAA). In this talk, I
will present some of the challenges and potential solutions for exa-scale
interconnection networks, which are being considered by IAA.
Back to Session III
|
|
Deployment Experiences,
Performance Observations, and Early Science Results on Ranger
John (Jay) R. Boisseau, Ph.D.
Director, Texas Advanced Computing Center
The University of Texas at Austin
The Texas
Advanced Computing Center (TACC) at The University of
Texas at Austin has deployed the largest open-science
supercomputing system
in the world. Ranger is a 1/2 petaflop Sun Constellation Cluster comprising
15,744 AMD Quad-core Opteron processors and a new Sun InfiniBand
interconnect, and uses a fully open-source software stack. The challenges of
deploying a Linux cluster of unprecedented scale with next-generation
components have been numerous, and the experiences gained in resolving then
should have great value in the HPC community. As the system is being
optimized, understanding of performance and scalability characteristics is
increasing, usage is growing, and some researchers are already computing at
32K processing cores with excellent scalability. This talk will summarize
the deployment experience, performance characteristics, and early science
results of Ranger and discuss some future challenges and implications.
Back to Session IV
|
|
NCSA Blue Waters: Preparing for
the Sustained Petascale System
Robert Pennington, National Center for Supercomputing Applications
University of Illinois at Urbana - Champaign
Urbana, IL, U.S.A.
The NCSA
Blue Waters system will be installed at the University of Illinois for production use in 2011. To prepare for this, there is significant
work to be done on software, applications, education/outreach as well as a
new building that will be constructed to house the system. The talk will be overview of the plans and
timeline for preparing for the system and a preliminary, high level summary
of the system capabilities.
Back to Session IV
|
|
The Impact of Petacomputing on
Models and Theories
Thomas Lippert
John von Neumann-Institute
for Computing (NIC)
FZ Jülich, Germany
In 2008,
supercomputers have reached the Petaflop/s performance level. Machines likes
the IBM Blue Gene/P, the Los Alamos Roadrunner or the IBM Ranger at TACC
achieve their unprecedented power using O(100.000) cores. In my talk I will,
on the one hand, discuss the question if we have arrived at the limits of
scalability – I will present first scalability results from the Jülich Blue
Gene/P system with 64k cores –, and, on the other hand, argue how
Petacomputers with hundreds of thousands of processors might transform
science itself.
Back to Session IV
|
|
Scalable Middleware for Large Scale Systems
Barton P. Miller
Computer Sciences Department
University of Wisconsin
Madison, Wisconsin, U.S.A.
I will discuss the problem of developing tools for large scale parallel
environments. We are especially interested in systems, both leadership class parallel
computers and clusters that have 10,000's or even millions of processors. The
infrastructure that we have developed to address this problem is called
MRNet, the Multicast/Reduction Network. MRNet's approach to scale is to
structure control and data flow in a tree-based overlay network (TBON) that
allows for efficient request distribution and flexible data reductions.
The second part of this talk will present an overview of the MRNet
design, architecture, and computational model and then discuss several of the
applications of MRNet. The
applications include scalable automated performance analysis in Paradyn, a
vision clustering application and, most recently, an effort to develop our
first petascale tool, STAT, a scalable stack trace analyzer running currently
on 100,000's of processors.
I will conclude with a brief description of a new fault tolerance
design that leverages natural redundancies in the tree structure to provide
recovery without checkpoints or message logging.
Back to Session IV
|
|
Old problems never die – managing
the multi-programming mix
Miron Livny
Computer Sciences Department
University of Wisconsin – Madison, WI, U.S.A.
Old
problems never die; they just fade away as technologies and tradeoffs
change. As the state of the art in
hardware and applications evolves further, they resurface. When virtual memory was introduced almost
50 years ago, computer systems had to find a way to prevent thrashing by
controlling the number and properties of the applications allowed to share
their physical memory. The recent proliferation of multi-core processors,
usage of virtual machines and deployment of complex I/O sub-systems require
the development of similar capabilities to control and manage at several
scales the mix of applications that share the compute and storage resources
of today’s systems.
Back to Session V
|
|
Active Data: Blurring the
distinction between data and computation
Tim Ho and David Abramson
Monash University
Clayton, Vic, Australia
The
amount of data being captured, generated, replicated and archived
on the Grid is growing at an astonishing rate; some researchers predict a
tenfold increase every five years from 2000 to 2015 alone. Managing this
growth will be challenging, and will require significant effort in the areas
of data management
and curation.
While it may now be feasible to store everything forever, it is wasteful and
we should not assume that we can continue doing this for very long. In
addition, much of the data requires active management, so that it can be
discovered, understood and reused by scientists and researchers within and
across disciplines. Clearly, there is a need to balance and justify the costs,
benefits and the anticipated needs
of future generation when considering maintaining large, distributed
archives indefinitely.
Fortunately, some datasets, such as those created by computations, are
reproducible. These datasets are typically derived from other data through a
process of computation, as opposed to those captured by instruments that are
often non-reproducible. Derived datasets can be reproduced by re-running the
computations that previously created them. Clearly, this requires sufficient
information on how data is computed, and a set of mechanisms that can create
it on demand.
This talk concerns problems that arise in the management and curation of
derived data. Specifically, we have developed a data lifecycle that models
the whole replication process, and, based on this lifecycle, designed and
prototyped an Active Data System that provides a complete, automated solution
to the above problems. In particular, our system allows users to recompute
data rather than necessarily store it, and adds a layer that provides
efficient access to remotely distributed replicated sources across different
middleware stacks.
Back to Session V
|
|
Using
Peer-to-Peer Dynamic Querying in Grid Information Services
Domenico
Talia and Paolo Trunfio
DEIS, University of Calabria
Rende, Italy
Dynamic
querying (DQ) is a technique adopted in unstructured Peer-to-Peer (P2P)
networks to minimize the number of nodes that is necessary to visit to obtain
the desired number of results. In this talk we describe the use of the DQ
technique over a distributed hash table (DHT) to implement a scalable Grid
information service. The DQ-DHT (dynamic querying over a distributed hash
table) algorithm has been designed to perform DQ-like searches over DHT-based
networks. The aim of DQ-DHT is two-fold: allowing arbitrary queries to be
performed in structured P2P networks, and providing dynamic adaptation of
search according to the popularity of resources to be located.
Through
the use of the DQ-DHT technique it is possible to implement a scalable Grid
information service supporting both structured search and execution of
arbitraries queries for searching Grid resources on the basis of complex
criteria or semantic features.
Back to Session V
|
|
Algorithms and scheduling techniques for
clusters and grids
Yves
Robert
Ecole
Normale Supérieure de Lyon, France
In this talk we provide several
examples to
illustrate key algorithmic concepts required to efficiently
execute applications on clusters and grids.
The idea is to give a lively exposition of the necessity to
inject whatever static knowledge is available into the design
of typical applications, such as master-slave tasking,
numerical kernels, and job workflows. We claim that this is the
key to an efficient deployment of these applications onto
large-scale distributed computational platforms.
The talk will proceed through examples to explain how to cope
with resource selection, memory constraints, platform heterogeneity, etc.
Back to Session V
|
|
Feedback
control for efficient autonomic solutions on the Grid
Rizos Sakellariou
University of Manchester
Manchester, United Kingdom
This talk
will consider different approaches for
autonomic solutions on the Grid. The talk will argue for the
need to use autonomic techniques to improve performance on
volatile environments such as those typically associated with
Grids and large-scale distributed computing. Examples will be
presented drawn from our work in the context of a UK and a
European research project.
Further information about the speaker: http://www.cs.man.ac.uk/~rizos
Back to Session V
|
|
Accidentally Using Grid Services
Charlie
Catlett
Maths
and Computer Science Division
Argonne
National Laboratory
Argonne, IL
and
University of Chicago
Chicago, IL, U.S.A.
Though
the term "grid" has fallen from the front page headlines, there is
an extremely active market of "grid services" - based on web
services and other standards - emerging. The web originally empowered
Internet users to create services and products with very little infrastructure,
and signs of success a decade ago included server meltdown from high demand.
Today one need not own any infrastructure at all to launch a new
service or product, and the combination of virtual and web services offers
not only near unlimited scaling but also reliability. This talk will
focus on a number of examples of new services, illustrating that at least one
measure of success is not only "ease of use" but "accidental
use" of transparent, but foundational, services.
Back to Session V
|
|
From Grid Computing to Cloud Computing
The evolution of the Grid Marketplace
Avner
Algom
The
Israeli Association of Grid Technologies
Israel
Over the last few years we have seen grid computing
evolve from a niche technology associated with scientific and technical
computing, into a business-innovating technology that is driving increased
commercial adoption. Grid deployments accelerate application performance,
improve productivity and collaboration, and optimize the resiliency of the IT
infrastructure.
Today, the maturity of the Virtualization
technologies, both at the VM and at the IT infrastructure levels, and the
convergence of the Grid, Virtualization and SOA concepts, enables the business
implementation of the Cloud Computing for utility and SaaS services.
At last, the Grid Computing vision becomes a
reality: people that get electricity from their electrical outlet, on-demand,
can get applications, computing and storage services from the network,
on-demand. We can dynamically scale our computation and storage power, at no
time, and we pay only for what we use.
This is going to change the marketplace as we know
it.
Back to Session V
|
|
Cloud Computing
for on-Demand Resource Provisioning
Ignacio Llorente
Distributed Systems
Architecture Group
Universidad
Complutense de Madrid
Madrid, Spain
The aim of the presentation is to show the
benefits of the separation of resource provisioning from job execution
management in different deployment scenarios. Within an organization, the
incorporation of a new virtualization layer under existing Cluster and HPC
middleware stacks decouples the execution of the computing services from the
physical infrastructure. The dynamic execution of working nodes, on virtual
resources supported by virtual machine managers such as the OpenNEbula
Virtual Infrastructure Engine, provides multiple benefits, such as cluster consolidation,
cluster partitioning and heterogeneous workload execution. When the computing
platform is part of a Grid Infrastructure, this approach additionally
provides generic execution support, allowing Grid sites to dynamically adapt
to changing VO demands, so overcoming many of the obstacles for Grid
adoption.
The previous scenario can
be modified so the computing services are executed on a remote virtual
infrastructure. This is the resource provision paradigm implemented by some
commercial and scientific infrastructure Cloud Computing solutions, such as
Globus VWS or Amazon EC2, which provide remote interfaces for control and
monitoring of virtual resources. In this way a computing platform could scale
out using resources provided on-demand by a provider, so supplementing local
physical computing services to satisfy peak or unusual demands. Cloud
interfaces can also provide support for the federation of virtualization
infrastructures, so allowing virtual machine managers to access resources
from remote resources providers or Cloud systems in order to meet fluctuating
demands. The OpenNEbula Virtual Infrastructure Engine is being enhanced to
access on-demand resources from EC2 and Globus-based clouds. This scenario is
being studied in the context of the RESERVOIR– Resources and Services
Virtualization without Barriers — EU-funded initiative.
Back to Session V
|
|
Network resource reservation and
virtualization for grid applications
Marcelo Pasin
INRIA, École Normale Supérieure
de Lyon
Laboratoire de
l’informatique du parallélisme
Lyon, France
The
coordination of grid resource allocation often needs a service to
transfer large datasets from one site to another within a
specified
time interval. Generally the transfer can start from any time after
its request and use any time variant bandwidth, as long as it is
completed before its deadline. We present BDTS, a new service for
Bulk Data Transfer Scheduling in the Grid, which provides users,
applications or grid middleware a service to specify transfer jobs and
ensure transparent control of them. It divides the active window of
a
transfer job into multiple intervals and independently assigns
bandwidth values for each of them producing bandwidth profiles. It
builds dynamic virtual networks by combining network bandwidth
reservation and on-demand optical link provisioning. In this
presentation, we discuss the conceptual foundations of the problem,
some implementation issues, and we present some experimental results.
Back to Session VI
|
|
A Performance
Based Distribution Algorithm for Grid Computing
Heterogeneous Tasks
Abderezak
Touzene, Hussein AlMaqbali, Ahmed AlKindi, Khaled Day
Department
of Computer Science Sultan Qaboos University
AL-Khod, Oman
Recently
in [1] we proposed a performance based load-balancing algorithm for
independent tasks, which require similar computing need in the sense that the
tasks are
almost identical. This paper extends the work and proposes a load
distribution algorithm for independent tasks with different computing
requirements including short and long tasks. We assume a preprocessing phase
of prediction of the number of instruction (TNI) needed for each task in the
grid. Our load distribution algorithm takes into account both the CPU speed
of the computing units and the TNI of different tasks. We design a simulation
model using steady-state, based on NS2 to study the performance of our load
distribution algorithm.
Keywords: grid
computing, load-balancing, steady-state, resource management, performance
evaluation, simulation models.
Back to Session VI
|
|
Parallel Data Mining from
Multicore to Cloudy Grids
Geoffrey Fox
Indiana University
Bloomington, IN, U.S.A.
We
describe a suite of data mining tools that cover clustering, Gaussian
modeling and dimensional reduction and embedding. These are applied to three
class of applications; Geographical information systems, cheminformatics and
bioinformatics. The data vary in
dimension from low (2), high (thousands) to undefined (sequences with
dissimilarities but not vectors defined). We use deterministic annealing to
provide more robust algorithms that are relatively insensitive to local
minima. We use embedding algorithms both to associate vectors with sequences
and to map high dimensional data to low dimensions for visualization. We
discuss the algorithm structure and their mapping to parallel architectures
of different types and look at the performance of the algorithms on three
classes of system; multicore, cluster and Grid using a MapReduce style
algorithm. Each approach is suitable in different application scenarios.
Back to Session VII
|
|
Summary-based Distributed Semantic Database for Resource and Service
Discovery
André Höing
Electrical Engineering and
Computing Science
Technical University of Berlin
Berlin, Germany
Today's
RDF triple stores that are based on distributed hash tables (DHTs) distribute
the knowledge of all participating peers in the P2P network. They use hash
values of the subject, predicate, and object of each triple in order to
identify three nodes in the network that shall store a copy of the triple.
Query processors collect relevant triples by identifying responsible nodes
using the hash values of literals and constants occurring in the query.
Usually, a soft state update mechanism is used to discard obsolete
information. This requires a periodic re-dissemination of all data which
creates significant network traffic despite the network being a bottleneck of
such systems. When considering resource and service discovery with highly
dynamic data like the system load, it is expensive and difficult to maintain
up-to-date information in the database.
This presentation introduces two novel ideas about managing semantically
enriched data in a DHT Peer-to-Peer network without such an enormous
overhead. We assume that in most cases RDF documents describing services and
resources have a structure that does not change much over time. Furthermore,
descriptions of similar instances also have a similar structure.
The first idea uses annotation of triples to label specific triples as
dynamic triples, where literals are substituted by pointers to locations
where dynamic data is stored.
The second idea goes one step further. Inspired by the idea of XML Data
Guides, we can build summaries of RDF documents that represent only the
structure but not the actual content of the documents. Such summaries are
smaller and more stable. By distributing the summaries in the network while
keeping the original data local to their source, the network overhead is
largely reduced. Most updates can be performed locally at a document's source
and do not generate any network traffic.
Before collecting the needed information for query processing, the query
algorithm must now identify nodes and documents that contain desired
information utilizing the distributed summary. This step can produce false
positives depending on the degree of detail in the summary. A main challenge
is to keep the number of false positives and the amount of network traffic as small as possible which
requires a sensible balance of the degree of details and the size of
summaries. Once nodes with relevant information have been determined, the
query processor collects
this information and processes the query thereon.
Back to Session VII
|
|
UNICORE 6 – A European Grid Technology
Achim Streit
Jülich Supercomputing
Centre (JSC) at Forschungszentrum
Juelich, Germany
The
development of UNICORE started back in 1997 with two projects funded by the
German ministry of education and research (BMBF). UNICORE is a vertically
integrated Grid middleware, which provides a seamless, secure, and intuitive
access to distributed resources and data and provides components on all
levels of a Grid architecture from an easy-to-use graphical client down to
the interfaces to the Grid resources. Furthermore, UNICORE has a strong
support for workflows while security is established through X.509
certificates. Since 2002 UNICORE is continuously improved to mature
production ready quality and enhanced with more functionalities in several
European projects. Today UNICORE is used in several national and
international Grid infrastructures like D-Grid and DEISA and is also
providing access to the national Supercomputer of the NIC in Germany.
The talk
will give details about the new version of UNICORE 6, which is web-services
enabled, OGSA-based and standards-compliant. To begin with the underlying
design principles and concepts of UNICORE are presented. A detailed
architecture diagram shows the different components of UNICORE 6 and its
interdependencies with a special focus on workflows. This is followed by a
view on the adoption of common open standards in UNICORE 6, which allows
interoperability with other Grid technologies and a realisation of an open
and extensible architecture. The talk closes with some interesting use case
examples, where the UNICORE Grid technology is used.
The
European UNICORE Grid Middleware is available as Open Source from http://www.unicore.eu.
Back to Session VII
|
|
e-Science Applications on Grids -
The DEISA Success Story
Wolfgang Gentzsch
DEISA Distributed European
Infrastructure for Supercomputing Applications
and
Duke University
Durham, North Carolina, U.S.A.
We will present
selected compute and data intensive applications which
have been ported to the DEISA Distributed European Infrastructure for
Supercomputing Applications. DEISA connects 11 powerful supercomputers in
Europe into a supercomputing grid
offering secure, transparent and remote
access to supercomputing cycles to every scientist in Europe. The
presentation will include lessons learned and recommendations for
applications to be ported on grids.
Bio:
Wolfgang Gentzsch
DEISA, Duke University
Wolfgang Gentzsch is Dissemination Advisor for the
DEISA Distributed European Initiative for Supercomputing Applications. He is
adjunct professor of computer science at Duke University in Durham, and
visiting scientist at RENCI Renaissance Computing Institute at UNC Chapel
Hill, both in North Carolina. From
2005 to 2007, he was the Chairman of the German D-Grid Initiative. Recently,
he was Vice Chair of the e-Infrastructure Reflection Group e-IRG; Area
Director of Major Grid Projects of the OGF Open Grid Forum Steering Group;
and he is a member of the US President's Council of Advisors for Science and
Technology (PCAST-NIT). Before, he was Managing Director of MCNC Grid and
Data Center Services in North Carolina; Sun's
Senior Director of Grid Computing in Menlo
Park, CA;
President, CEO, and CTO of start-up companies Genias and Gridware, and
professor of mathematics and computer science at the University of Applied
Sciences in Regensburg, Germany.
Wolfgang Gentzsch studied mathematics and physics at the Technical Universities
in Aachen and Darmstadt, Germany.
Back to Session VII
|
|
Superlink-online
- delivering the power of GPUs, clusters and opportunistic grids to
geneticists
M. Silberstein
Technion-Israel Institute of Technology
Haifa, Israel
Genetic linkage analysis is a statistical tool used
by geneticists for mapping disease-susceptibility genes in the study of
genetic diseases. The analysis is based on the exact inference in very large probabilistic
(Bayesian) networks, which is often computationally hard (ranging from
seconds to years on a single CPU).
We
present a distributed system for faster analysis of genetic data, called
Superlink-online. The system achieves high performance through parallel
execution of linkage analysis tasks over thousands of computational resources
residing in multiple opportunistic computing environments, aka Grids. It
utilizes the resources in many available grids, unifying thousands CPUs over
campus grids in the Technion and the University
of Wisconsin in
Madison,
EGEE, Open Science Grid and Community Computing Grid Superlink@Technion.
Notably,
the system is available online, which allows geneticists to perform
computationally intensive analyses with no need for either
installation of software, or maintenance of a complicated distributed
environment. It is being extensively used by medical centers worldwide,
running over 15,000 interactive genetic analysis tasks since 2006 and
consuming about 360 CPU years in all grids.
While the grids potentially provide enormous amount
of computing power, we also explore an alternative approach of using Graphics
Processing Units (GPUs) to accelerate the genetic linkage computations. We
achieve up to two orders of magnitude speedups on average, and up to three
order of magnitude speedups on some particularly complex problem instances
versus the optimized application performance on a single CPU. The use of GPUs
is particularly appealing in the context of Community Grids, considering
the number of high performance GPUs available worldwide.
In this talk we describe various aspects of the system architecture
which drives Superlink-online, including scheduling, resource allocation,
fault tolerance and reliability, as well as our recent GPU-related results.
Back to Session VII
|
|
Building Collaborative Applications for System-Level Science
Marian Bubak
Institute of Computer Science AGH, al. Mickiewicza 30,
30-059 Krakow, Poland
ACC CYFRONET AGH, Krakow, ul.
Nawojki 11,
30-950 Krakow, Poland
A novel
approach to scientific investigations, besides analysis of individual
phenomena, integrates different, interdisciplinary sources of knowledge about
a complex system to obtain an understanding of the system as a whole. This
innovative way of research has recently been called system-level science [1].
Problem-solving
environments and virtual laboratories have been the subject of research and
development for many years [2]. Most of them are built on top of workflow
systems [3]. Their main drawbacks include limited expressiveness of the
programming model and lack of mechanisms for integration of computing
resources from grids, clusters and dedicated computers.
The
ViroLab project [4] is developing a virtual laboratory [5] for research of
infectious diseases to facilitate medical knowledge discovery and provide
decision support for HIV drug resistance [6], and this virtual laboratory may
be useful in
other
areas of system-level science.
To
overcome the limitations of the programming methods, we have defined an
experiment plan notation based on a high-level scripting language - Ruby. For easy interfacing of
different technologies, we have introduced a grid object abstraction level
hierarchy [7]. Each grid object class is an abstract entity which defines the
operations that can be invoked from the script, each class may have multiple
implementations, representing the same functionality; and an implementation
may have multiple instances,running on different resources [8].
The
Experiment Planning Environment is an Eclipse-based tool supporting rapid
experiment plan development while Experiment Management Interface enables
loading and execution of experiments. The Experiment Repository stores
experiment plans prepared by developers and published for future usage, and
the laboratory database holds the obtained results.To enable high-level
programming, the virtual laboratory engine, called the GridSpace, includes
the Grid Operation Invoker which instantiates grid object representatives and
handles remote operation invocations. The GridSpace Application Optimizer is
responsible for optimal load
balancing
on computational servers.The Data Access Service acquires data from remote
databases located in research institutions and hospitals. To meet the
specific requirements for exchanging biomedical information within such a
virtual environment, the solution introduced in DAS bases on existing Grid
technologies: Globus Toolkit, OGSA-DAI, and Shibboleth. The provenance
approach [9] in the ViroLab virtual laboratory brings together ontology-based
semantic modeling, monitoring of applications and the runtime infrastructure,
and database technologies, in order to collect rich information concerning
the execution of experiments, represent it in a meaningful way, and store it
in a scalable repository [10].
The
virtual laboratory has already been used to plan and execute a few
virological experiments, with various types of analysis of HIV virus
genotypes such as calculation of drug resistance based on virus genotype,
querying historical and provenance information about experiments, a drug
resistance system based on the Retrogram set of rules, data mining and
classification with Weka [5], and the
molecular
dynamics NAMD application which has been installed on the CYFRONET EGEE site.
The
virtual laboratory provides an environment to collaboratively plan, develop
and use collaborative applications; it is dedicated for multi-expertise
task-oriented groups running complex computer simulations; its basic features
are: mechanisms for user-friendly experiment creation and execution,
possibility of reusing existing
libraries,
tools etc., gathering and exposing provenance information, integration of
geographically-distributed data resources, access to WS, WSRF, MOCCA
components and jobs, secure access to data and applications.
Acknowledgments
The
Virtual Laboratory is being developed at the Institute of Computer
Science
and CYFRONET AGH, Gridwise Technologies, Universiteit van Amsterdam,
and HLRS
Stuttgart in the framework of the EU IST ViroLab and CoreGRID
projects
as well as the related Polish SPUB-M and Foundation for Polish
Science
grants.
References
[1] I. Foster and C. Kesselman: Scaling
System-Level Science: Scientific
Exploration and IT Implications,
IEEE Computer, vol. 39, no 11, 31-39,
2006
[2] K. Rycerz, M. Bubak, P.M.A. Sloot, V.
Getov: Problem
Solving Environment for Distributed
Interactive Simulations in: Sergiei
Gorlatch, Marian Bubak, and Thierry
Priol (Eds). Achievements in European
Reseach on Grid Systems. CoreGRID
Integration Workshop 2006 (Selected
Papers) ISBN-13: 978-0-387-72811-7; pp
55 - 66, Springer, 2008
[3] Y. Gil, E. Deelman, M. Ellisman, T. Fahringer, G. Fox,
D. Gannon, C. Goble, M. Livny, L. Moreau, and
J. Myers.
Examining the Challenges of Scientific
Workflows. IEEE
Computer vol
40, no 12 pp 24-32, 2007
[4] ViroLab - EU
IST STREP Project 027446; www.virolab.org
[5] ViroLab Virtual
Laboratory, http://virolab.cyfronet.pl
[6] P. M.A. Sloot,
I. Altintas, M. Bubak, Ch.A. Boucher:
From Molecule to Man: Decision
Support in Individualized E-Health,
IEEE Computer vol. 39, no 11, 40-46,
2006
[7] T. Gubala, M. Bubak: GridSpace -
Semantic Programming Environment
for the Grid, PPAM'2005, LNCS 3911,
172-179, 2006
[8] M. Malawski, M. Bubak, M. Placek, D.
Kurzyniec, V. Sunderam:
Experiments with Distributed Component
Computing Across Grid Boundaries,
Proc. HPC-GECO/CompFrame Workshop -
HPDC'2006, Paris, 2006
[9] D. de Roure, N.R. Jennings, N. Shadbolt,
The semantic grid:
a future e-science infrastructure, Grid
Computing -
Making the Global Infrastructure a
Reality,
Wiley, 2003, pp. 437-470
[10] B.
Balis, M. Bubak, and J. Wach: User-Oriented Querying
over Repositories of Data and
Provenance,
In G. Fox, K. Chiu, and R. Buyya,
editors, Third IEEE International
Conference on e-Science and Grid
Computing, e-Science 2007,
Bangalore, India, 10-13 December 2007, pages
77-84.
IEEE Computer Society, 2007
Back to Session VII
|
|
DMOVER: Scheduled Data
Transfer for HPC Grid Workflows
Derek Simmel
Pittsburgh Supercomputing Center
Pittsburgh, PA, U.S.A.
TeraGrid
users have expressed a need for better tools to schedule and
manage data transfer among distributed HPC resources and instruments
involved in computational workflows. Users require assured access to data
transfer services and reliable means to coordinate data transfers
before,
during, and after computational jobs and instrument data collection
events. The interface to these coordinated data transfer
services must be
easy to use, consistent and interoperable among different HPC resources,
instruments, and computational workflow tools. Scheduled data transfer
tasks must be reliable, repeatable, and exhibit
consistent performance
for similar transfer parameters and conditions.
DMOVER addresses these needs using familiar job scheduling methods to
allocate and prepare WAN-connected nodes dedicated to data transfer
tasks. DMOVER is a portable harness used to schedule and automate data
transfers via Globus GridFTP and GSI-SCP. Users submit DMOVER jobs using
the dsub command-line client, with applicable parameters and a task
file
containing a list of transfer source an destination URLs. When a DMOVER
job starts, it automatically configures and deploys
optimized GridFTP and
GSI-SSH servers on the allocated nodes. Transfers specified in the
tasks
file are then executed, supported by automated
retries as well as user
credential monitoring and renewal. Detailed logs are collected to
facilitate monitoring of data transfer tasks.
Coordinated workflows
involving computational and DMOVER jobs are
implemented via compatible
job managers or Globus job submission services.
Back to Session VII
|
|
Grid Computing or the Internet of services?
Opportunities and perspectives from research to business
Antonio Congiusta
NICE-ITALY, Cortanze,
Asti, Italy
Experience has shown that solutions to
better enable organizations to take advantage of the benefits of Grid
computing, are based on clear identification of the requirements and the
application of the best available standardized and reliable technologies.
Relevant examples of such principle with
related best practices can be extracted from some of the success stories that
recently have involved EnginFrame in the Oil & Gas industry, the Energy
and Automotive sectors, HPC support from collaboration facilities to
infrastructure provision and management, and also some fruitful cooperations
with strategical partners.
In particular, beyond to well
established HPC activities within a primary European consortium for providing
a production quality infrastructure, a new trend has been undertaken towards
the integration of collaboration facilities to HPC environments. Quite
interesting are also the activities devoted to enable for workflow management
and distributed visualization, some of which are part of European-wide
research projects.
From all such experiences we can
envision as future of the Grid an always strong evolution towards interoperable
key services, within a scenario in which comprehensive and all-inclusive
software is ever less important. In such a scenario, a key role is played by
integration technologies capable of homogenizing and enforcing service
interactions and access.
Back to Session VII
|
|
e-Research
& Grid computing in Australia: From Infrastructure to Research
David Abramson
Monash University
Clayton, Vic, Australia
Over the past
few years the Australian government has performed a major review of its
research infrastructure needs, from hard technological areas to the
social sciences. Along with this review, they have investigated the
electronic platforms required to support these various disciplines.
What has evolved is an grid computing strategy called "Platforms
for Collaboration" that addresses computation, networking and data
management. In addition to this, various computer science groups are developing
grid technologies that underpin this platform. In this talk I will give
an over of the Australian e-Research agenda and highlight a few major
research activities in grid computing.
Back to Session VIII
|
|
Grid and e-Science in Korea
Kihyeon
Cho
e-Science Division
Korea Institute of Science and Technology Information Daejeon,
305-806, Korea
For Grid and e-Science in Korea we had been
focused on the research of Grid technology and Grid infrastructure till 2006.
Since 2007, we have been in the stage of enabling science on the cyber
infrastructure for four major sciences such as e-Life Science, e-Physics,
e-Engineering, and e-Geo Science. In order to
support these application areas we also work on scientific workflow technology
and WSRF (Web Service Resource Framework) based Grid service technology. We
also provide researcher collaborations with both infrastructure and
technology. In this talk, we will present the
current status of Grid and e-Science projects and activities in Korea.
Back to Session VIII
|
|
Grid Activity in India
Atul Gurtu
Tata Institute of Fundamental
Research
Mumbai, INDIA
Grid technology has changed the way advanced
research is being conducted today. In India too,
the main driver for introduction of GRID activity has been participation in
High Energy Physics (HEP) experiments at the LHC. To extract meaningful
physics results within a reasonable span of time from peta-bytes of data of
unprecedented complexity and to include effective participation of
collaborating institutions spread out world wide, the only way was to develop
GRID based distributed computing. Integration of various distributed
environments in different administrative domains having varied security
policies poses new challenges in data security & data sharing. The
required computational performance is supported through world wide LHC
Computing Grid (WLCG). The progress and status of setting up WLCG networking
and computing in India will be
described with main emphasis on the High Energy Physics related activity.
Status of Tier-2 centers for participation in the ALICE and CMS experiments
at the LHC at CERN will be presented. The role of the EU-India GRID project
in developing Grid based solutions in other areas of science such as earth sciences, bio, condensed matter physics etc. will be
mentioned as also some domestic GRID initiatives within the country.
Back to Session VIII
|
|
National Grid Initiative of Ukraine
Anatoly Sachenko
American-Ukrainian School of Computer Science
Department of Information
Computing Systems and Control
Ternopil State Economic University
Ternopil, Ukraine
Uniting of
the existing Grid segments and supercomputer centers in scientific and
educational areas into joint
Ukrainian National Grid Initiative(UNGI) and the issues of UNGI integration
into the European Grid infrastructure are considered in this paper. The
peculiarities of Grid segment at
National Academy of Science as well as the UGrid Project of Ministry of
Education and Science are described too. It’s stressed on the joint project
UNGI for EGI and other integration possibilities within INTAS, NATO and Frame
7 programs. Finally an advanced approach for security strengthening in Grid-systems is proposed.
Back to Session VIII
|
|
The European Grid Initiative
Per Öster
CSC – Finnish IT Center for
Science
Espoo, Finland
The European Grid Initiative (EGI) has as goal to ensure
a long-term sustainability of grid infrastructures in Europe. This is to be done through establishment of
a new federated model bringing together National Grid Infrastructures to
build the EGI Organisation. For this purpose, the European Commission has
funded a specific project, the EGI Design Study (EGI_DS), to in 27 months
make the conceptual setup and operation of a new organisational model of a
sustainable pan-European grid infrastructure. The goal of the EGI Design
Study (EGI_DS) is to evaluate use cases for the applicability of a
coordinated effort, to identify processes and mechanisms for establishing
EGI, to define the structure of a corresponding body, and ultimately to
initiate the construction of the EGI Organisation. The project started in
September 2007 and a very important milestone is the Blueprint of the EGI
Organisation. In this talk the EGI Blueprint and feedback from the NGIs will
be presented together with a discussion of the role of EGI in the European
“HPC Ecosystem”.
Back to Session VIII
|
|
Building Shared High Performance
Computing Infrastructure for the Biomedical Sciences
Marcos Athanasoulis, Dr.PH, MPH
Harvard Medical School
U.S.A.
In recent years high performance
computing has moved from the sidelines to the mainstream of biomedical
research. Increasingly researchers are employing computational methods
to facilitate their wet lab research. Some emerging laboratories and
approaches are based on a 100% computational ramework. While there are
many lessons to be learned from the computational infrastructure put into
place for the physical and mechanical sciences, the character, nature and
demands of biomedical computing differ from the needs of the other sciences.
Biomedical computational problems, for example, tend to be less
computationally intensive but more “bursty” in their needs. This
creates both an opportunity (it is easier to meet capacity needs) and a
challenge (job scheduling rules are more complicated to accommodate the
bursts). Harvard Medical School provides one of the most advanced shared
high performance research computing centers at an academic medical center. In
2007, Harvard convened the first Biomedical High Performance Computing Leadership
Summit to explore the issues in creating shared computing infrastructure for
the biomedical sciences. We brought together over 100 leaders in the
field to exchange ideas and approaches. Through special sessions and
direct participant surveys a number of themes emerged around best practices
in deploying shared computational infrastructure for the biomedical
sciences. Based on prior experience and the summit findings, we
summarizes obstacle and opportunities facing those who wish to provide biomedical
oriented high performance computing infrastructure. We will provide
quantitative results about HPC Biomedical Implementations. We will also
provide some examples of current problems in biomedical computations
including whole genome processing, massively parallel correlation analysis
and natural language processing of clinical notes.
Back to Session IX
|
|
ViroLab: Distributed Decision Support in a virtual laboratory for
infectious diseases
P. Sloot
University of Amsterdam
Amsterdam, Netherlands
In future
years, genetic information is expected to become increasingly significant in
many areas of medicine. This expectation comes from the recent and
anticipated achievements in genomics, which provide an unparalleled
opportunity to advance the understanding of the role of genetic factors in
human health and disease, to allow more precise definition of the non-genetic
factors involved, and to apply this
insight rapidly to the prevention, diagnosis and treatment of diseases.
Virolab integrates the biomedical information from viruses (proteins and
mutations), patients (e.g. viral load) and literature (drug resistance
experiments) resulting in a rule-based distributed decision support system
for drug ranking. In addition ViroLab includes advanced tools for (bio)
statistical analysis, visualization, modelling and simulation, enabling
prediction of the temporal virological and immunological response of viruses
with complex mutation patterns for drug therapy.
Back to Session IX
|
|
Processing of Large-Scale Biomedical Images on a Cluster of Multi-Core
CPUs and GPUs
Umit Catalyurek
Department of Biomedical
Informatics
The Ohio State University
Columbus, Ohio, U.S.A.
As
microprocessor manufacturers strain to continue to increase performance,
multi-core chips are quickly becoming the norm. The demand in computer gaming
industry also brought us GPUs as an alternative fast, general purpose,
streaming co-processors.
Commodity
GPUs and multi-core CPUs bring together an unprecedented
combination
of high performance at low cost, and provide an ideal environment for
biomedical image analysis applications.
In this
talk we will present our ongoing efforts on developing optimized biomedical
image analysis kernels for heterogeneous multi-core CPUs and GPUs. We will
also present how a cooperative cluster of multi-Core CPUs and GPUs can be
efficiently used for large scale biomedical image analysis.
Back to Session IX
|
|
Grid
Computing for Financial Applications
M. Al-Baali§,
P. Beraldi*, L. Grandinetti*, G. Aloisio^ I.
Epicoco^, A. Violi**, C. Figŕ Talamancaç
§ Dept. of Mathematics and Statistics,
Sultan Qaboos University, Muscat, Oman
* Department of Electronics,
Informatics and Systems, University of Calabria
* * CESIC - University of Calabria
^ University of Salento
ç Innova spa
In recent
years financial operators have shown an increasing interest in quantitative
tools able to efficiently measure, control and manage risk. Such an interest
is motivated by the necessity to operate in a very competitive and volatile
environment with a high level of
complexity increased by the
globalization of the economic activities and the continuous
introduction of innovative financial products. The complexity of the problems
to deal with and the necessity to operate in real time has highlighted the
serious computational constraints imposed by conventional numerical
platforms, prompting the need to take advantage of high performance computing
systems.
In this
talk we present a prototypal system designed to support financial operators
in investment decisions concerning the strategic asset allocation
problem. The system has been designed
and tested within the European Project BEINGRID.
At the
core of the system is the formulation of sophisticated optimization models
able to capture with an increasing level of realism with respect to
traditional approaches, the specific features of the applicative problem.
Moreover, the system is based on the integration of advanced scenario
generation procedures and efficient methods to solve the resulting huge sized
problems.
The
system has been deployed on the SPACI grid infrastructure. In particular, an
user – friendly web grid environment
has been realized by using the GRB technology for the resource management and
the GRelC services for distributed data.
Back to Session IX
|
|
Data Issues in a challenging HPC
application to Climate Change
Giovanni Aloisio
University of
Salento
Lecce, Italy
Earth
Science is strongly becoming a data intensive and oriented activity.
Petabytes of data, big collections, huge datasets are continuously produced,
managed and stored as well as accessed, transferred and analyzed by several
scientists and researchers at multiple sites. From the data grid perspective,
a key element to search, discover, manage and access huge amount of data
stored within distributed storages is the related data and metadata
framework. A new supercomputing centre, the Euro-Mediterranean Centre for
Climate Change (CMCC), was recently created by the Italian Government to support
research on Climate Change. The SPACI Consortium, one of the main CMCC
Associate Centres, provides know-how and expertise on High Performance and
Grid Computing. The GRelC Middleware (provided by SPACI Consortium) has been
recently adopted as part of the CMCC Data Grid framework in order to provide
a secure, transparent and scalable grid enabled metadata management solution.
We
present the CMCC initiative, the supercomputing facility as well as data grid
architectural and infrastructural issues concerning the adopted grid
data/metadata handling systems.
Back to Session IX
|
|
Tim David
Centre for Bioengineering University of Canterbury
Christchurch, New Zealand
“A
Heterogeneous Computing Model for a Grand Challenge Problem”
Back to Session IX
|
|
The e-Science for High Energy Physics
Kihyeon Cho, Ph.D.
KISTI (Korea Institute of Science
and Technology Information)
Daejon, Korea
The
e-Science for High Energy Physics is to study High Energy Physics (HEP) any
time and anywhere even if we are not on-site of accelerator laboratories. The
components are 1) data production, 2) data processing and 3) data analysis
any time and anywhere. The data production is to do remote control and take
shifts remotely. The data processing is to run jobs anytime, anywhere using
Grid farms. The data analysis is to work together to publish papers using
collaborative environment. We apply this concept to LHC experiment at CERN
and Tevatron experiment at Fermilab. It this talk we will present the current
status and embodiment of the idea.
Back to Session IX
|
|
A HPC infrastructure at the service of Scientific Research in Italy
Giovanni
Erbacci
CINECA -
System and Technology Deparment
CINECA Inter-University Consortium,
Casalecchio di Reno, Italy
State of the art HPC infrastructures are fundamental
to support scientific research and to advance science at European level.
Since many years, at Italian level, CINECA has been able to assure to the
scientific community a competitive
advantage by putting into timely production advanced HPC systems that have proven very wide applicability
and success.
The CINECA HPC infrastructure de facto represents
the national facility for supercomputing and the CINECA HPC systems are part of the Italian research
Infrastructures system, integrated by means of the Italian academic and
research network facility (GARR).
In this work we present the CINECA HPC
infrastructure, its evolution, and the
service model. Moreover, we outline
the CINECA role in the context of the
main HPC Infrastructure projects, operating
at the European level: DEISA, PRACE and HPC-Europa.
DEISA is a consortium between the most advanced HPC centres in Europe, whose
aim is to deploy and operate a persistent, production quality, distributed
supercomputing environment with continental scope.
This infrastructure is mainly intended to support
challenge scientific applications by integrating and making easily accessible
supercomputers in different centres.
PRACE is a feasibility project intended to build
the next generation of challenge HPC infrastructure and services at European level. The infrastructure will
consist of a limited number (3 to 5) of
PetaFlop/s class HPC systems integrated in a network of HPC systems on
a pyramidal model basis, with three different layers (European, National and Regional) in the European HPC eco-system.
HPC-Europa supports the human network of
knowledge, experiences and expertise exchange, in the context of the
scientific research communities using
advanced HPC systems.
HPC-Europa factively promotes such mission
supporting the mobility of the European researchers among the main research
institutions, and providing the access
to the computational resources offered by the main European HPC
infrastructures.
Back to Session IX
|
Main Aim 
