October 10th-14th 2005
Following the success of the 1st Grid Plugtests, during the 10th-14th of October 2005 the 2nd Grid Plugtests was held. Organized by ETSI and INRIA, the objectives were: to test Grid interoperability, and to learn, through the user experience and open discussion, about the future features needed for Grid middlewares.
The 2nd Grid Plugtests consisted of several events: Conferences, Workshops, Tutorials and a Contest. Drawing over 240 participants from many different countries. The events were organized as follows:
These events were organized by ETSI Plugtests and the INRIA OASIS research team. OASIS is a joint team between INRIA, UNSA, I3S-CNRS which develops the ProActive Grid middleware, hosted by ObjectWeb. The event was officially sponsored by e-Europe, IBM, Microsoft, SUN Microsystems, and financially supported by Region PACA, INRIA, I3S. The Flowshop contest was sponsored by ROADREF.
To run experiments on Grid computing, a Grid was setup for three days with the help of numerous partners. This Grid was deployed on 13 different countries, in more than 40 sites, gathering 2700 processors for a grand total of more than 450GFlops (measured with the SciMark 2.0 benchmark).
Given the heterogeneity of the sites, each site had to be configured and fine tuned. This involved figuring out the operating system, installing an adequate Java Virtual Machine for the operating system (when not already installed), figuring out the network/firewall configuration, job scheduler, etc. This worked was handled by the OASIS Team, mainly Romain Quilici and Mario Leyton, who prepared the Grid for the contest and Plugtests.
The deployment was thus made very simple and transparent for the Plugtests users, who had all the architecture details hidden by the ProActive layer.
ProActive is a LGPL Java library for parallel, distributed, and concurrent computing, also featuring mobility and security in a uniform framework. With a reduced set of simple primitives, ProActive provides a comprehensive API allowing to simplify the programming of applications that are distributed on Local Area Networks (LAN), on clusters of workstations, or on Internet Grids.
The deployment descriptors provide a mean to abstract from the source code of the application any reference to software or hardware configuration. It also provides an integrated mechanism to specify external process that must be launched, and the way to do it. The goal is to be able to deploy an application anywhere without having to change the source code, all the necessary information being stored in an XML Deployment Descriptor file.
Since programming the Grid cannot be achieved at a low-level of abstraction, ProActive is provided with a programming model. The complexity that arises from scale, heterogeneity, and dynamicity cannot be tackled with message-level primitives. As such, development of new Grid programming models have to rely on higher-level of abstraction that the current usage. These programming models are based on the component technology.
The following steps describes, in a broadly manner, the methodology used to configure each site for the Grid. Average time of configuration varied depending on the complexity of the site from less than an hour to several weeks.
Steps 3 and 4 were the most demanding for the OASIS Team, since they required careful inspection of the site, and sometimes protocol interoperability development (see Sections , ).
Steps 5 and 6 were fairly easy to build, and proved to be most useful during the Plugtests. On the first place, to install (when it was requested by the contestant) the application libraries (jar). This was done to improve the deployment time by avoiding dynamic class loading. Secondly, for cleaning the Grid between each contestant's run.
Steps 7 and 8 were fairly simple when not dealing with complex configuration sites, but when facing problems usually required to go back and fine tune a previous step until eventually the site was correctly configured.
Figuring out the environment configuration of a site was a key process in building the Grid. Given the heterogeneousness of the Grid, the environment varied considerably from site to site. The most important aspects of the environment can be grouped into the following areas: Operating System & JVM, Schedulers and Site Access, Network and Firewalls and Data Storage.
Since ProActive requires Java to operate, a very important aspect of site configuration is to determine whether a JVM is installed on the site, and therefore on each node of the Grid. On some cases, after searching in the remote sites, a proper JVM was found to be installed and was used.
When no JVM was found, the Operating System (Linux, AIX, SGIIrix, MacOS, Solaris) and hardware architecture had to be determined (x86_32, x86_64, ia64, AIX, SGIIrix, PPC, Sparc). Afterwards, a suitable JVM had to be installed, preferably the Sun version, but if not suitable, then an alternative was used (IBM, Apple, Sgi). To avoid requesting privileged permits on the site, the JVM installation took place on the home directory of the site account.
We can classify the access into two categories depending on the scheduler/middleware installed on the site: remote or local access. Remote access is used with deployment protocols such as Globus, Unicore, NorduGrid, GLite where the job submission takes place directly from a client machine, usually with a certificate scheme provided by the protocol.
On the other hand local access protocols like: LSF, PBS, OAR, PRUN are used locally at the site, and therefore an SSH (or equivalent) connection must be combined with the job submission protocol. With ProActive, this can be easily done using the Deployment Descriptor. Nevertheless, to avoid password prompts an ssh passphrased key was installed on the remote sites to allow non interactive access using an ssh agent.
The network can be classified into different levels of security policies.
The data storage scheme varied from site to site. On many of them, the Network File System (NFS) was used, thus sharing the home user directory overall nodes on the site. These cases were the most simple to configure, since the software installation (ProActive and JVM if necessary), only had to take place once. On the other hand sites which did not share the user home directory proved to be very troublesome, specially for configuring the synchronization scripts.
One difference from last year with respect with data storage, was that some new protocols like NorduGrid or Unicore provide the concept of Job Space. When a job is submitted using any of this protocols, a specific space for the job is created on the cluster. This job space is temporal, and can be used by the process to store data. Nevertheless when the process finishes, the Job Space is destroyed. Thus making a persistence installation difficult. To solve this issue, our approach was to use Deployment File Transfer (see Section ).
Support for several new deployment protocols were developed. This was necessary to include new partners into the Grid. Also, several new features were added to ProActive to cope with specific site configurations like Hierarchical Deployment and File Transfer.
Among the new deployment protocols that were developed to interface with other middlewares or schedulers we can find: OarGrid, NorduGrid, Unicore and GLite.
Hierarchical Deployment was a key feature developed for this years Grid Plugtests. Following from last years experience, many sites had configurations that used internal IP networks, or were located behind a very restrictive firewall. During the 1st Grid Plugtests, it was up to the user to provide a forwarding mechanism for accessing the internal Grid nodes. Since this proves to be very complicated at the user application level, and taking last year's Plugtests experience into account, this year the OASIS Team, mainly Clement Mathieu, Romain Quilici and Matthieu Morel, worked on providing transparent support at the ProActive level for inner site nodes. As a result, sites could be added to the Grid requiring less configuration effort by the site's administrators.
Nevertheless this feature is still in a development status, with many improvements and bug fixes pending. For example, during the Plugtests one of the teams realized that the Group feature can not be combined, at this point, with Hierarchical Deployment. Thus, the Plugtests experience provided important feedback for ProActive improvements.
Another interesting feature that was developed corresponds to Deployment File Transfer support. This allows the user to specify files that need to be transfered at deployment time to the Grid nodes. The main result of this approach is that ProActive can be transfered on-the-fly along with the JVM on sites which do not allow persistant software installation (a job space is created for each submitted job, and later destroyed when the job is finished). Sites that used this mechanism were NorduGrid, Unicore and GLite.
For the 2nd Grid Plugtests, more than 40 sites located on 13 different countries were configured by the OASIS Team. The complexity of configuring each site varied, as described in Section .
Here we present the list of sites that formed part of the Grid. To easy the readability, we have sorted this sites alphabetically first, and secondly by site name. For this same reason we have also grouped them into four Tables: , , , . The columns of each table are described as:
Figure shows a graphical representation of the Grid. The location of sites are pointed out with flags in the map. This maps shows how we reached a worldwide dissemination, with sites in Asia, Australia, Europe, North and South America. The details for each site can be found in Tables: , , , .
To benchmark the Grid we used the SciMark agent for computing. This measure was taken using a pure Java benchmark. Since the types of JVMs used were heterogeneous in vendor and version, comparing Mflops between sites is pointless. More over, given the instability of a Grid of this nature (in size and location), for some sites we were unable to obtain all the provided resources at the moment of benchmark. In this cases, we extrapolated to estimate the total site capacity. Because all of this reasons, the specified Grid benchmark is a very rough one, and should be considered as a reference, and not a certification or a rigorous scientific study.
Considering the 1st Grid Plugtests benchmark (100GFlops), this years corresponds to a significant improvement: 450GFlops (approx). The details of this computation can be found in Tables: , , , .
Figure shows the distribution of number of CPUs per Grid site. Figure shows the Distribution of Grid Processing capacity (in Mflops). The graph in Figure holds both results in a pie like graph.
As last year, many difficulties were encountered when setting up the Grid.
For example, when dealing with Grid5000 we faced some problems with the oardel command on some sites. The command did not execute correctly a job deletion, and we had to write a script to do this. Other problems we faced were oarsub not working correctly with parameter ``all'' on the Grid500 Sophia site, and other small details. Nevertheless, the monitoring webpage provided by Grid5000 proved to be very useful to diagnose and solve the problems.
Also on Grid5000, we developed the support for the oargrid submission protocol, which was finally not used (we had to fall back on the oar protocol), because the oargridsub command provided a very rigid behaviour: exactly all the requested nodes were returned for all sites, or no nodes were returned at all. When dealing with requests for hundreds of nodes, it is very likely that some might fail. For us, it would have been much more useful if the oargridsub command provided more flexibility by allowing the specification of ``at most'' or ``at least''.
On other Grids we also faced some problems. For CNGrid we had to implement a custom LSF protocol for one of the sites. Also, the provided resource for CNGrid were very busy (not exclusive access for the Plugtests), and most of the time we were unsuccessful at submitting a large job reservation.
For Unicore we developed our own interface using the testing server. Unfortunately, we were unable to test our interface when dealing with big sites, since Unicore only provided virtual sites with one processor.
With the GLite interface we faced some problems when trying to deploy from a Fedora Core 3 (FC3) machine. We discovered at this point that the GLite client library is not supported for FC3 and newer. We managed to solve this problem by using ProActive's remote job submission features. To achieve this, we deployed the job submission first into a GLite client compatible machine using ssh, and from there submitted the job to the GLite machines. The transfering of the GLite JDL file was handled using ProActive's File Transfer mechanism.
Even though hierarchical deployment was a key feature for this Plugtests, it still lacks maturity and development for more complex scenarios. We would like to continue development this feature, since we believe is fundamental for building the Grid.
Finally, we had some problems with the Unix/Linux open file (connection) limitation. For a Grid of this size the default limitation on current distributions is 1024 . This is too small when we take into account that this years Grid involved over 2700 processors. ProActive provides a mean to reduce the number of open connection by specifying that this should be closed in the deployment descriptor files. None the less, this optimization was not enough for a Grid of this size. We therefore incremented the default value to (16K) for the contests machines. Nevertheless, we only realized during the Plugtests that hierarchical deployment provided an even harder stress on the open file limits. For example, we had to contact the administrator to increment this limitation for the Grid500 Sophia site.
Overall these difficulties proved to be a valuable part of the Grid interoperability, and will help us to continue improving and developing the Grid.
This year, two contest were organized during the 2nd Grid Plugtests. Like the last year, the N-Queens Counting Problem was present: How many ways can N queens be placed on a NxN chessboard. Also, a new problem was added this year, the Flowshop Problem: What is the optimum way of using M machines for J jobs were a job j in a machine m takes P time.
These events were strictly an engineering event, not a conference, nor a workshop. As such, an active participation was requested from the companies/organizations which had to write their own implementation of the problem. There was no compulsory programming language, all teams used Java, and when possible, some used native code inside a Java wrapper.
Four teams competed this year in the N-Queens contests. The criterion for deciding the winners were based on:
The FlowShop contest was sponsored by ROADREF providing a prize of 500 Euros. Each team had 1 hour to run their application on the Grid. During this period, they were expected to solve Taillard's instances of the FlowShop problem. The instances were required to be solved exactly with proof of optimality. This means that the program must find the exact solution, and prove that it is the optimal. If more than one team solved the problem correctly, the winner was the one that solved the problem in less elapsed time. If more than one team solved the same problem in the same amount of time, the final criteria for deciding the winner was the number of workers (number of CPUs) used.
For the contests and tutorial, 25 machines were installed and configured by the OASIS Team. The main software installed on the machines were: Fedora Core 3, Sun JDK1.4.9, Eclipse, ProActive and other contest environment configuration. One of the machines was configured as a central server for the user accounts using NFS. In order of arrival to the ETSI Plugtests room, each team was assigned an account on the machines, from team1 to teamN. Contestants spent the first day (and part of the second) testing and fine tuning their code for the Grid.
Florian Martin, from the OASIS Team, worked on preparing the remote contest participation at Santiago. Using the time zone difference, the remote contest took place mainly during the night, which corresponded to the afternoon in Santiago, allocating exclusive access to the Grid during this period.
Basically Florian Martin's job was to contact Grid actors in south America, negotiate access and configure them into the Grid. He also had to organize the Plugtest in Santiago, to allow the local teams to participate to the event. For this, a special room was reserved for the event. Each participant used an individual local computer, and each machine was connected to one of the ETSI contest machines, thus allowing them access to the Grid.
These results are taken from the ETSI 2nd Grid Plugtests N-Queens Challenge Results report. The contests results are as follows:
These results are taken from the ETSI 2nd Grid Plugtests Flowshop Challenge Results report. The results are detailed as follows:
After the Plugtests the N-Queens Challenge was extended for one month. This gave an opportunity for the motivated teams to continue testing Grid operability. The remote challenge results are shown in Figures , , and.
Using the same criteria as for the Plugtests N-Queens Challenge, the results were as follows:
Table shows a comparison chart of the 1st and 2nd Grid Plugtests. This differences have been mentioned through this report, but are summarized here. From the table, it is easy to see that the 2nd Grid Plugtests embraced an even wider range of the Grid community.
The Grid gathered for the 2nd Grid Plugtests proved to be heterogeneous in many levels: Computer Arquitecture, Operating Systems, Java Virual Machines, Deployment Protocols and Network Configurations. The diversity of resources is detailed as follows:
The 2nd Grid Plugtests, co-organized by INRIA and ETSI, pleased all the participants. It was an event useful for the Grid community: users and developers. The Conferences and Workshops helped the community to exchange their views, objectives, difficulties and user experience for developing the Grid. Also, with the Tutorials, the gap between the users and the Grid was narrowed by presenting the different Grid tools and middlewares.
In the specific case of ProActive, the Plugtests gave us the opportunity to develop new and interesting features, while testing the middleware at a new level of complexity. The results shown during the N-Queens and Flowshop contests left us very happy, since they showed that the applications could take advantage of the heterogeneous Grid in a simple way.
As usual, setting up the Grid proved to be a lot of hard work with problems and difficulties. The OASIS Team had to implement new deployment protocols, and new ways to adapt to network configurations. This new tools were an important advancement from last year, since they enabled more restrictive sites to join the Grid with less effort from the sites administrators. Nevertheless, after the Plugtests experience we believe these tools still require further development before they can become an integral feature of ProActive. The Plug & Play Grid is still not a reality, but after the Plugtests we can happily say that it lies one step closer.
Given the positive experience of the event, we would like to organize a 3rd version. In this occasion, we would like to encourage a wider usage palette of tools for accessing and programing the Grid. We would also like to have a wider community involvement, including new organizations, for example, GGF and EGA.
This document is taken from the on-line version .
Australia UNIVERSITY OF MELBOURNE Rajkumar Buyya <email@example.com.OZ.AU> Srikumar Venugopal <firstname.lastname@example.org.OZ.AU> Brazil LNCC Bruno Schulze <email@example.com> Chile DCC Universidad de Chile Jose Piquer <firstname.lastname@example.org>, Florian Martin <Florian.Martin@sophia.inria.fr>, Luis Mateu <email@example.com> Chile UTFSM Xavier Bonnaire <firstname.lastname@example.org> China BUPT MA Yan <email@example.com> Xiaohong Huang <firstname.lastname@example.org> China CNGRID Zhang Xiaoming <email@example.com> China CNGRID-ICT Zhang Xiaoming <firstname.lastname@example.org> China CNGRID-NHPCC Zheng Fang <email@example.com>, <firstname.lastname@example.org> China CNGRID-HKU Lin Chen <email@example.com> China CNGRID-SCCAS Zhang Xiaoming <firstname.lastname@example.org>, Sungen Den <email@example.com> China CNGRID-SCCNET Jiang Kai <firstname.lastname@example.org> China CNGRID-USTC PengZhan Liu <email@example.com> France IDRIS-DEISA Victor Alessandrini <firstname.lastname@example.org>, Philippe Collinet <email@example.com>, Gilles Gallot <Gilles.Gallot@idris.fr> France INRIA SOPHIA-ANTIPOLIS Nicolas Niclausse <Nicolas.Niclausse@sophia.inria.fr>, Francis Montagnac <Francis.Montagnac@sophia.inria.fr>, Janet Bertot <Janet.Bertot@sophia.inria.fr>, Jean-Luc Szpyrka <Jean-Luc.Szpyrka@sophia.inria.fr>, Antoine Zogia <Antoine.Zogia@sophia.inria.fr>, Regis Daubin <Regis.Daubin@sophia.inria.fr> France GRID5000-BORDEAUX Aurelien Dumez <firstname.lastname@example.org> France GRID5000-GRENOBLE Nicolas Capit <<email@example.com> France GRID5000-LYON Frederic Desprez <firstname.lastname@example.org>, Stephane D'Alu <email@example.com> France GRID5000-ORSAY Philippe Marty <firstname.lastname@example.org>, Gilles Gallot France GRID5000-RENNES Guillaume Mornet <email@example.com>, David Margery <David.Margery@irisa.fr> France GRID5000-SOPHIA Sebastien Georget <Sebastien.Georget@sophia.inria.fr>, Nicolas Niclausse <Nicolas.Niclausse@sophia.inria.fr> France GRID5000-TOULOUSE Celine Juan <firstname.lastname@example.org>, Pierrette Barbaresco <email@example.com> France LIFL Melab Nouredine <Nouredine.Melab@lifl.fr>, El-ghazali Talbi <El-ghazali.Talbi@lifl.fr>, Sebastien Cahon <Sebastien.Cahon@lifl.fr> France LORIA Xavier Cavin <Xavier.Cavin@loria.fr>, Bertrand Wallrich <Bertrand.Wallrich@loria.fr>, Alain Filbois <Alain.Filbois@loria.fr>, Olivier Demengeon <firstname.lastname@example.org>, Benjamin Dexheimer <Benjamin.Dexheimer@loria.fr> France SUPELEC Stephane Vialle <email@example.com>, Patrick Mercier <Patrick.Mercier@supelec.fr> Germany UNICORE Daniel Mallmann <firstname.lastname@example.org> Greece FORTH ICS Manolis Marazakis <email@example.com> Ireland QUEEN'S UNIVERSITY OF BELFAST Ron Perrott <firstname.lastname@example.org>, Andrew Carson <a.carson@Queens-Belfast.AC.UK> Italy BENEVENTO Eugenio Zimeo <email@example.com>, Nadia Ranaldo <firstname.lastname@example.org> Italy ISTI Domenico Laforenza <email@example.com>, Ranieri Baraglia <Ranieri.firstname.lastname@example.org>, Giancarlo Bartoli <email@example.com> Italy UNIVERSITY OF PISA Marco Danelutto <firstname.lastname@example.org>, Pietro Vitale <email@example.com> Netherland VRIEJ UNIVERISTY Kees Verstoep <firstname.lastname@example.org>, Henri Bal <email@example.com>, Norway NORDUGRID Oxana Smirnova, Aleksandr Konstantinov, Balazs Konya, Alexander Lincoln Read, <firstname.lastname@example.org> Switzerland CERN/GILDA TESBED (Italy) Bob Jones <Robert.Jones@cern.ch>, Marc Ozonne <Marc.Ozonne@sophia.inria.fr>, Roberto Barbera <email@example.com>, Giuseppe Platania <firstname.lastname@example.org> Switzerland EIF Pierre Kuonen <email@example.com>, Jean-Francois Roche <firstname.lastname@example.org> Switzerland ETHZ Luc Girardin <email@example.com> USA UC IRVINE Stephen Jenks <firstname.lastname@example.org> USA USC - CENTER FOR GRID TECHNOLOGIES Mats Rynge <email@example.com>
Informations that we need to know about machines you are going to provide for the Grid Plugtests:
This document has been taken from the online version.
Two contests will take place during the Plugtests: The N-Queens Counting Problem and The Flowshop Problem. To solve these problems, a world wide Grid will be configured, composed of a rich diversity of systems (architecture, operating system and Java virtual machines).
The following document is intended to help contestants fine tune their applications to compete at the Plugtests event. Grid Architecture
The Grid will be composed of more than a 1000 CPUs. These CPUs will be distributed all around the world, and grouped into sites. The size of all sites will be heterogeneous, ranging from a handful to hundreds.
To deploy on each site, contestants will use already configured Deployment Descriptors. There will be one deployment descriptor per site, configured with a virtualnode named "plugtest". The name of this virtualnode is the one that should be hard-coded into the contestant's application code. The length of the node array (number of nodes) returned from the virtualnode will vary depending on the size of the site.
The machines on a site may have one or more CPUs. For each site with more than one CPU the configuration can be one of the following:
It is highly discouraged to use static variables in the deployed active objects. Sometimes, more than one active object will be deployed on the same same java virtual machine, which may produce a race conflict between the active objects.
Last year experience shows this is a latent risk, and must be avoided.
Native code, is highly discouraged. The first reason is the heterogeneousness of the Grid, since code will require specific compilation for each site. The second reason is that size of the Grid, which makes it unfeasible to compile and copy the native code to the remote sites during the plugtest event. The third and last reason not to use native code is that by using it your team will limit the amount of machines to which it can deploy, reducing the Grid capacity.
The ProActive / Plugtests staff will not provide support for native code during the plugtest event.
The machines of a site can have: private or public IPs. For sites with private IPs, ProActive has provided a new feature that will allow deployment between the site's frontend and the innermachines.
Nevertheless, the current status of this feature does not support inner node communication between two different sites. That is to say, if site A and site B have inner nodes: A1...AN, B1...BM, then Ax will not be able to communicate with By.
For security reasons, solutions which require communication between tasks will be limited to a subset of the sites known as Grid5000 (composed of more than a thousand CPUs).
Due to the large number of descriptor files, the deployment time is significant. Therefore it is recommended to contestants to deploy each descriptor file in parallel thread.
Moreover, the process of placing active objects on the nodes can also be done in parallel, using a thread pool.
Teams are expected to:
This document was generated using the LaTeX2HTML translator Version 2002-2-1 (1.70)
Copyright © 1993, 1994, 1995, 1996,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.
The command line arguments were:
latex2html -split 1 report.tex
The translation was initiated by Mario Leyton on 2006-01-13