The Netherlands present themselves at SC|07 in Reno NV with a demonstration of several innovative and interesting projects. This is a continuation of the Dutch presence in Baltimore MD (2002), Phoenix AZ (2003), Pittsburgh PA (2004), Seattle WA (2005) and Tampa FL (2006).
The "Dutch Pavillion" is cooperation and combined show of major research institutes and not-for-profit research network infrastructure and supercomputing organisations and funding agencies.
What can be seen?
BiG Grid
Project
Color-Based Object
Recognition by a Grid-Connected Robot
Dog
Jave-Centric Grid
Computing
LOFAR: the eyes of
the World
Network
Description Language
Personal Space
Station
SCARIe
project
Signal Detection
in Climate Projections
StarPlane: An
Application Controlled Photonic
Network
Surfnet
Introduction
BiG Grid is a large scale effort to realise a major grid for Science and Research in The Netherlands. The project has become known under the name "BiG Grid", and has been targeted to a broad range of scientific research disciplines which can take advantage of the ICT resources, available in a science grid infrastructure.
While the Netherlands has been a leading player in the development of the grid, and has considerable expertise in bio-informatics, distributed sensors networks, and particle physics, the large-scale infrastructure to fully exploit this position of leadership is missing. It has been the purpose of the BiG Grid proposal to realise such a science-wide grid infrastructure in the Netherlands.
A basic ingredient for the proposed infrastructure is the network. The Netherlands are already in an excellent position, due to the world-class network services provided by SURFnet. The currently available resources (like the National Supercomputer, compute cluster capacity, data storage facilities, etc.) will need to be "gridified" and extended to a level which enables ground-breaking scientific research.
The BiG Grid project is funded for the period 2007-2012 with a total budget of M€ 28.8. The BiG Grid project will cover a full range of hardware entities, but also a significant amount of the budget is reserved for manpower, to attract and to bring a wide range of scientific disciplines to the science grid. The project has been initiated by the Netherlands National Computing Facilities Foundation (NCF), the National institute for subatomic physics (Nikhef) and the Netherlands Bio-Informatics Centre (NBIC).
Mission
The project's mission is:
To realise a fully operational world-class and resources-rich grid environment at the national level in the Netherlands to serve public scientific research, including particle physics, life sciences and all other disciplines, and to encourage actively general grid usage across all disciplines.
The science case for the BiG Grid proposal is the integral of many different science cases, reflecting the broad scientific community base. The realisation of BiG Grid is crucial to the success and continuity of many Dutch research groups, covering important areas such as life sciences in a broad context, astronomy, particle physics, meteorology and climate research, water management, to name but a few. The very nature of the new infrastructure, a multidimensional collaboration enabler and accelerator, allows for direct participation from social sciences, humanities, and also addresses communities in administrative domains (administrative grids such as digital academic repositories).
The realisation of the grid infrastructure provides opportunities for enhanced international visibility. Dutch participation in international generic grid developments is already prominent (in flagship projects like EGEE and DEISA) and are on a national scale covered by the BSIK funded VL-e project (Virtual Laboratory for e-Science), whilst for the life sciences, coordinated by the Netherlands Genomics Initiative, NBIC (Netherlands BioInformatics Centre), partly BSIK funded as well, is the key player for enabling bio-informatics methodology.
Realisation of this major project puts the Netherlands at the forefront of grid developments. The infrastructure enables many national ambitions. The excellent position of Dutch academic hospitals in their collections of patient data can be enhanced by using the grid for biobanking. Major advances in drug discovery are enabled through combining data from various research communities as well as through the availability of massive compute resources enabling direct drug interaction modelling. The infrastructure allows industrial research labs (e.g. Philips) to profit from the available resources. It allows the LOFAR project to not only position itself as a multi sensor radio-astronomy centre but also as the European centre from which a variety of scientific communities using the LOFAR data will be served. It positions the Netherlands as one of the (worldwide only) ten Tier-1 sites for CERN's LHC experiments.
More information
More information can be obtained from the BiG Grid project office or from the BiG Grid website (www.biggrid.nl).
BiG Grid Project Office p/a NWO/NCF
Postbus 93575
2509 AN Den Haag
michielse@nwo.nl
F.J. Seinstra and J.M. Geusebroek
ISLA, Informatics Institute, University of Amsterdam
Kruislaan 403, 1098 SJ Amsterdam, The Netherlands
{fjseins, mark}@science.uva.nl
Abstract
Multimedia data is rapidly gaining importance along with recent developments such as the increasing deployment of surveillance cameras in public locations. In a few years time, analyzing the content of multimedia data will be a problem of phenomenal proportions, as digital video may produce data at rates beyond 100 Mb/s, and multimedia archives steadily run into Petabytes of storage space. Consequently, for urgent problems in multimedia content analysis, Grid computing is rapidly becoming indispensable.
This demonstration shows the viability of wide-area Grid systems in adhering to the heavy demands of a real-time task in multimedia content analysis. Specifically, we show the application of a Sony Aibo robot dog, capable of recognizing objects from a set of learned objects, while connected to a large-scale Grid system comprising of cluster computers located in Europe, the United States, and Australia. As such, we demonstrate the effective integration of state-of-the-art results from two largely distinct research fields: multimedia content analysis and Grid computing.
1. Introduction
Irrespective of the application of a robot, the problem of object recognition is to determine which, if any, of a given repository of objects appears in an image or video stream. It is a computationally demanding problem that involves a non-trivial tradeoff between specificity of recognition (e.g., discriminating between different faces) and invariance (e.g., to different lighting conditions). Due to the rapid increase in the size of multimedia repositories consisting of 'known' objects [3], state-of-the-art sequential computers no longer can adhere to the computational demands, making high-performance distributed computing indispensable.
This demonstration shows a real-time object recognition task performed by a Sony Aibo robot dog, connected to a Grid system spanning our entire globe (see Figure 1). In the first 'learning' phase of the demonstration, we present a set of objects under a single visual setting. In the second 'recognition' phase, we validate the learning step
by showing the objects again, under varying lighting conditions, lighting color, and viewing position. In earlier experiments we have shown that our dog is capable of accurately recognizing more than 300 objects from a set of 1,000 learned objects, under a diversity of imaging conditions. Interestingly, this recognition rate is higher than the recognition rate of around 200 objects reported for a real dog [4].
Figure 1: Object recognition by our robot dog: (1) an object is held in front of the dog's camera; (2) video frames are processed on a per-cluster basis; (3) given the resulting feature vectors describing the scene, a database of known objects is searched; (4) in case of recognition, the dog reacts accordingly (see also: www.science.uva.nl/~fjseins/aibo.html).
2. Color-Based Object Recognition
Color is a powerful cue in the recognition of objects. Recognition based on color, rather than just intensity, provides a broader class of discrimination between objects. The use of RGB values, however, does not directly increase recognition performance, certainly not when variations in imaging conditions are encountered. Differences in intensity, direction, and color of the illumination, as well as shading and cast-shadow significantly affect the appearance of an object. Therefore, it is meaningful to transform the RGB values to invariant properties, which relate to surface properties rather than to object appearance. We previously derived a broad class of invariants [1, 2], which are shown to be robust under noisy conditions. Furthermore, these invariants can be scaled to the size of the object structure. Recognition with these invariants boils down to learning an invariant representation of the object, rather than learning every possible appearance of a single view of the object.
We decompose the recognition of object appearance into two schemes. First, we have different views or aspects of an object, each of which has to be learned. Secondly, there is the illumination, drastically influencing object appearance. For this class of appearance effects, we demonstrate invariants to be very effective. Object recognition may be based on a weak description of the important features in the scene, as long as mutual correspondence between observation and objects in the world is maintained. Therefore, we learn local histograms of invariant features for each aspect of an object.
Figure 2: In the recognition phase, the dog will make a 360 degrees turn, recognizing the objects it has learned previously (see also: www.science.uva.nl/~fjseins/aibo.html).
3. Services-based Multimedia Grid Computing
From the user's perspective, Grid computing is still far from being more than just an academic concept. Essentially, this is because Grids do not yet have the full basic functionality needed for extensive use. Consequently, as long as programming and usage is hard, most researchers in multimedia computing will not regard Grids as a viable alternative to more traditional computer systems. The Parallel-Horus project [5, 6, 7] attempts to overcome this problem by providing a software architecture that allows multimedia researchers to implement fully sequential applications for efficient execution on cluster systems. To further stimulate the use of clustered wide-area Grid systems in the multimedia community, the user transparent programming model is supported by an easy-to-use execution model based on Wide-Area Multimedia Services, i.e. high-performance multimedia functionality that can be invoked from sequential applications running on a desktop machine. With Parallel-Horus, dynamic systems of distributed multimedia services, in which clients and servers can participate at will, can be created without any parallelization and distribution effort from the user. Our robot dog application indeed constitutes a dynamic system of this kind.
4. Demonstration
The color-based object recognition is performed using the video data obtained from the camera hidden in the dog's nose. Each video frame is processed on one of the available clusters (see Figure 1). Using this world-wide computing capacity, we first demonstrate the learning phase, by presenting our dog with new objects, potentially obtained from people in the audience. Next, we demonstrate the recognition phase, by letting our dog walk around in a circle and indicate the objects it has previously learned (see Figure 2).
Acknowledgements
The authors would like to thank all researchers and system maintainers who have granted us access to their cluster computers. We are also grateful to Michiel van Liempt for his excellent efforts in implementing the original sequential object recognition code. Thanks go out to Edwin Steffens and Arnoud Visser for providing us with a robot dog.
References
Jeroen van der Ham, Freek Dijkstra, Paola Grosso, Cees de Laat, University of Amsterdam
NDL web page: http://ndl.uva.netherlight.nl/
Computer networks form an important part of our daily lives. Networks are becoming faster and more complex. The Network Description Language (NDL) provides an interoperable way to describe computer networks. These descriptions allow applications to have a better understanding of the network so they can more easily adapt it to their needs, and also allow reducing the perceived complexity for the user.
The Network Description Language is developed by the University of Amsterdam and aims to describe computer networks in a meaningful way. Using the Resource Description Framework (RDF), NDL provides an ontology for computer networks. With this ontology you can create a concise description of a network.The need of a shared common vocabulary that describes networks is particularly important in optical networking, and specifically in facilitating the multi-domain provisioning of lightpaths.
Bandwidth on Demand
As optical networking grows into a mainstream technology, more users want to use a dedicated circuit through the network. For example businesses want their own optical private network (OPN) to connect their remote office locations with the main site, and scientists want to use lightpaths to transfer huge amounts of data over dedicated connections. NDL can help in the process of network provisioning: end users can unambiguously express requests for connections with two or more end-points. These requests can be submitted to the network provisioning software, and if the connection is inter-domain, the same request can be forwarded to other domains. NDL can be used to forward the request to other domains.
Network provisioning is especially hard for connections that traverses multiple domains, and this is precisely were NDL shines. Like pointers to services, in RDF you can also point to other network domains, paving the way for a distributed, thus scalable information database. Furthermore, the NDL domain schema allows network operators to automatically create abstracted views of the network, providing only that information to the neighboring domains which is strictly required.
Visualization
From NDL descriptions we can also easily create visualizations of the network. An advantage of RDF is that it allows descriptions to link to other descriptions. Using this in NDL means that inter-domain connections can be described using these links, while keeping the descriptions under control of their domains.
NDL allows large inter-domain networks, such as the Global Lambda Integrated Facility (GLIF) to easily create network maps. Network descriptions in NDL can use the geo RDF schema, describing the location of network devices. With this information, a map can automatically be created using Google Maps.
Path Finding
In the past pathfinding in optical networks was performed by engineers using pictures of networks. An important design goal of NDL is to provide enough information for pathfinding. NDL is currently used in production software at SARA, the Dutch super-computing center, to find available end-to-end SDH connections through the Dutch research network SURFnet 6.
Virtually all path finding software is only able to find network connections at a given layer. Path finding in a network with dynamics at multiple layers is inherently more complex. In fact, in certain circumstances, the shortest viable path through a multilayer network may contain loops. The University of Amsterdam has developed the Python NDL Toolkit that is able to not only describe these situations in NDL, but also find paths through such a network using a modified breadth first search algorithm.
NDL facilitates path finding in multi-layer networks by providing both ontologies to describe networks, as well as an ontology to describe network technologies. This feature is important for the multilayer path finding algorithm. Since it allows the creation of generic technology descriptions, the multilayer path finding algorithm does not have to have any knowledge about specific technologies, but only needs knowledge about the generic characteristics of the technologies. This means that there is no need to alter the algorithm as new technologies emerge.
Demonstrations
At the Dutch Pavilion at SC07 we have a demo stand with two NDL demonstrations. We demonstrate NDL-based multi-layer path-finding, as described in the previous paragraph: we show that shortest paths through multi-layer networks can contain loops.
The second demonstration is on creating network testbeds; using NDL as input, we emulate a network using User-Mode Linux. This allows engineers to test and experiment without affecting their production networks.
References
Abstract
The SCARIe project is a cooperation between JIVE, SARA and the University of Amsterdam aimed at building a distributed software correlator for real-time processing of astronomical VLBI data. By delivering high quality pictures of the deep sky, VLBI is a powerful tool for astronomers. The software correlator requires fast networking connectivity for handling the high data rates in real-time. Hence, the DAS-3 grid and its user-controllable optical network, Starplane, form an ideal platform for the software correlator.
VLBI and JIVE
Very Long Baseline Interferometry (VLBI) is a technique in astronomy in which several distant radio telescopes observe the same object simultaneously. By using VLBI, astronomers can make detailed images of cosmic radio sources with unsurpassed resolution. The data is processed by correlating each pair of input signals. JIVE operates a dedicated hardware correlator specifically designed to perform this task. Currently, the recorded data are sent from the telescopes to JIVE through mail. The next step, e-VLBI, uses the Internet to stream data in real time to the correlator. Besides connecting telescopes in real-time to the correlator, JIVE is conducting several projects (Fabrics, SCARIe) to investigate the use of a software correlator on a Grid infrastructure and to increase the flexibility and capacity of the correlation process.
What is SCARIe?
SCARIe aims at developing a software correlator operating on grids. A first prototype was built to track the Huygens probe during its descent on Titan, one of Saturn's moons, and convincingly demonstrated the flexibility of a software correlator. We estimate that 3 Tflops are needed to equal the capabilities of the hardware correlator. The challenge is not in the amount of "flops" but in the fact that this correlator is supposed to operate in real-time even for large amounts of data (7.2 TB per hour).
SCARIE on DAS-3/StarPlane
We are using the ASCI DAS-3 grid to develop and conduct our software correlation experiments. DAS-3 is composed of five cluster sites that are connected by a photonic network in which the StarPlane project is pioneering "application controlled photonic networks".
In distributed software correlation the incoming data are divided into small chunks that are correlated independently. Starplane permits us to dynamically adapt the network topology to the amount of data we have to transmit and we are developing transmission patterns that take profit of it. For example, by switching the lightpath in a circular manner it becomes possible to send data to all clusters at high throughput (up to 20Gb/s according to the StarPlane roadmap).
Conclusion and future plans
During the first year of the project we have laid the foundation for a flexible software correlator based on distributed computing technologies. Using this software correlator we are currently collaborating intensively with the StarPlane project in order to implement a scaleable software correlator on the whole DAS-3 grid.
Li Xu1, JP Velders1, Paola Grosso1, Jason Maassen2, Kees Verstoep2, Cees de Laat1, Henri Bal2
1: Universiteit van Amsterdam, Kruislaan 403, 1098SJ, Amsterdam, The Netherlands Tel: +31-20-525-7531, Fax: +31-20-525-7490 Email: {lixu, jpv, grosso, delaat}@science.uva.nl
2: Vrije Universiteit, De Boelelaan 1081A, 1081HV, Amsterdam, The Netherlands Tel: +31-20-598-7733, Fax: +31-20-598-7653 Email: {jason, versto, bal}@cs.vu.nl
Overview
In e-Science environment, applications increasingly demand more flexibility from the networks to satisfy their computational requirements; these requests can be accommodated on optical networks with the creation of dedicated lightpaths. What we still miss are flexible network infrastructures that enable applications to control the underlying lightpath topology at runtime.
The StarPlane research project aims at building a pioneering "application controlled photonic network". The project is funded by Dutch Research Council (NWO) and carried out by the researchers at the University of Amsterdam (UvA) and the Vrije University (VU). It enables applications running on the Grid cluster - Distributed ASCI Supercomputer 3 (DAS-3) to request lightpaths on a dedicated portion of the Dutch research and education network - SURFnet6.
Infrastructure
To create a shared computing infrastructure for their students and researchers a couple of Dutch universities that participate in the Advanced School for Computing and Imaging (ASCI) created DAS-3. DAS-3 is composed of five clusters with about 270 dual-CPU nodes supercomputers in total, which are integrated into a single large-scale distributed system. The five DAS-3 clusters are located in Leiden, Delft and in Amsterdam at VU and at UvA. The clusters installed at each location differ slightly in total number of nodes, type of processors, amount of storage and memory but they all share the same high-level architecture, as shown in Fig.1.
SURFnet6, entered production in 2006, deploys multiple fiber optic rings that connect the various academic and research locations around the Netherlands. One of these photonic rings connects the universities in Leiden, Delft and Amsterdam, the locations of the DAS-3 clusters. A portion of the wavelengths on the ring provides the infrastructure for StarPlane lightpaths. This network was built using Nortel Networks equipment which is a combination of the Common Photonic Layer (CPL), the Optical Multiservice Edge (OME) 6500 and the Wavelength Selective Switch (WSS). Fig.2 shows in more detail the DAS-3 optical wide-area setup.
Figure 1 DAS-3 cluster architecture
Figure 2 DAS-3 wide-area connection
Novelty
One novelty of the StarPlane is that it puts the control of the network directly in the hands of the application. In StarPlane the network is completely under the control of the applications provided they have the appropriate authorization levels. The network provider does not influence how the application can use the network, and does not provision the lightpath on behalf of the user. Multiple applications can run simultaneously on separate lightpaths while the StarPlane middleware orchestrate their operation.
A second innovative aspect of StarPlane is that application will be allowed to change the underlying photonic topology in real-time, the goal being the provisioning of lightpaths in the sub-seconds timescale. This requires not only the photonic hardware (e.g. WSS) in the network to react in this time frame to changes requests, but also the StarPlane middleware to handle and process the application request in this timescale.
Middleware design
The overall software architecture of StarPlane is composed of:
e-Science applications on StarPlane
In StarPlane we also investigate how applications can benefit from the flexibility and performance of this photonic network. Several classes of applications are being considered, ranging from file transfers, to eVLBI and medical modeling. For example, Awari, a communication-intensive distributed game-tree search application, achieves a performance on distributed clusters that is close to that of a single larger cluster.
Demo
In SuperComputing'07, we collaborate with people from SARA to demonstrate a monitoring system that shows the real-time StarPlane network utilization and functionalities of StarPlane Management Plane (SPMP) in the aspect of path reservation and provisioning. There are two levels of monitoring being performed: cluster-level and lightpath-level. In the cluster-level monitoring, we record the packet number on the Myrinet switches and display the dynamic summary of the StarPlane network. In the lightpath-level monitoring, we retrieve the statistic data from the OME6500 optical devices. The interactions with SPMP for lightpath reservation are also demonstrated.
More information
Network facilities for higher education and research. SURFnet enables breakthrough education and research. We develop and operate the hybrid SURFnet6 network and provide innovative services in the areas of security, authentication and authorisation, group communication and video.
SURFdiensten.
Every day SURFnet provides access to the Internet to over 750,000 scientists, teachers and students in higher education and research. This allows them to securely send large amounts of data and to communicate with other network users around the world.
Collaboration and innovation
SURFnet combines the demand of institutions for higher education and research and in doing so creates advantages of scale and collaboration for all connected institutions. Sustained innovation means that users always have one of the fastest and most advanced networks in the world at their disposal. Furthermore SURFnet enables multimedia collaboration between institutions, researchers and students through advanced middleware and applications. In this manner SURFnet provides the ICT foundation that underpins innovation in higher education and research in the Netherlands.
SURFnet stimulates the telecommunications and Internet market in the Netherlands. The new possibilities thus find their way to other sectors, reinforcing the knowledge economy in the Netherlands.
A high-grade and reliable network
SURFnet provides the national and international network facilities for the higher education and research community in the Netherlands. We interconnect the local networks of over 180 connected institutions and link them to the rest of the world. Our hybrid network is among the fastest and most innovative networks in the world.
GigaPort
The development of SURFnet6 takes place within the context of the GigaPort project. This five-year collaboration between public and private sectors started on 1 January 2004 under the name GigaPort Next Generation Network. GigaPort's goal is to reinforce the national knowledge infrastructure.
Secure and advanced Internet services
SURFnet continually develops new Internet services for institutions and end users:
SURFnet's pioneering role
For years the SURFnet network and services have been recognised as leading and ranking among the most advanced in the world. In its pioneering role SURFnet puts in a sustained effort to develop knowledge and experience on new technologies. SURFnet shares this knowledge and experience with its users as well as the international community.
In order to contribute to the development and standardisation of new technology SURFnet actively participates in international projects and organisations. These include the European GÉANT project and TERENA, the collaborative organisation of national research networks. SURFnet also maintains close links with international organisations such as Internet2, the American research network, and the Canadian research network, CANARIE. SURFnet arranges standardisation agreements in these international settings to enable the use of applications in cross-border collaborations.
SURFnet is part of SURF. SURF is the collaborative organisation for higher education institutions and research institutes aimed at breakthrough innovations in ICT. Other SURF-organisations are SURFfoundation and SURFdiensten.
