Grid Computing is an exciting buzzword in the computing world today. Here we define it to mean the exploitation of a varied set of networked computing resources, including large or small computers, PDAs, file servers and graphics devices. The networks could be anything from high speed ATM to wireless or modem connections. Exploiting these connected resources could, for example, enable large scale simulations not possible on a single supercomputer, aid computational work of geographically distributed collaborations, simplify remote use of machines, and enable the new dynamic application scenarios we propose. Two important aspects of Grid technology, which have been largely ignored, form the basis of our GridLab Project, which aims to build components for Grid applications (as MatLab does for mathematics), and realistic testbeds for their development:

1. Co-development of Infrastructure and Applications:

   We propose a balanced program with co-development of a range of Grid applications (based on Cactus, the leading, widely used Grid-enabled open source application framework, and Triana, a dataflow framework used in gravitational wave research) alongside infrastructure development, working on transatlantic testbeds of varied supercomputers and clusters. This practical approach ensures that the developed software truly enables easy and efficient use of Grid resources in a real environment. We will maintain and upgrade the testbeds through deployment of new infrastructure and large scale application technologies as they are developed. All deliverables will be immediately prototyped and continuously field tested by several user communities. Our focus on specific application frameworks allows us immediately to create working Grid applications to gain experience for more generic components developed during the project.

2. Dynamic Grid Computing:

   We will develop capabilities for simulation and visualization codes to be self aware of the changing Grid environment, and to be able to fully exploit dynamic resources for fundamentally new and innovative applications scenarios. For example, the applications themselves will possess the capability to migrate from site to site during the execution, both in whole or in part, to spawn related tasks, and to acquire/release additional resources demanded by both the changing availabilities of Grid resources, and the needs of the applications themselves.

Figure 1: GridLab architecture

We will unify these elements in developing innovative, practical, Grid computing technologies, which will then be quickly and easily adopted and exploited by applications from many different research and engineering fields, as shown in Figure 1. Specific key objectives of our project are to:

• Design and develop a Grid Application Toolkit (GAT), to provide core, easy to use functionality through a carefully constructed set of generic APIs for both simulation codes and Grid software. The GAT will contain independent modules for handling many different aspects of Grid programming, including simulation, performance and grid monitoring, resource brokering and selecting, performance prediction, interaction with information servers, security, notification, collaboration, data handling, remote visualization, and remote application steering.
• Simultaneously enhance real applications for the Grid, implementing new dynamic simulation scenarios using the GAT. Both Cactus and Triana will be extended to integrate and exploit GAT elements, making Grid Computing easily exploitable by a wide range of applications. Our simulation driven, compute intensive applications are fundamentally different from the highly data driven applications in many other Grid projects (e.g., DataGrid, GriPhyN, EuroGrid).
• Develop and test Grid infrastructure/applications on real testbeds, constructed by linking heterogeneous collections of supercomputers and other resources spanning Europe and the USA, using and extending existing testbeds. Interoperability with different testbeds will be ensured by also using production testbeds in the USA, driving international high speed network connectivity. Testing will be carried out by the project and by several large, closely related user communities, including an EU Astrophysics Network, and various multidisciplinary US funded collaborations.

User Scenarios

1. Gravitational Wave Detection and Analysis:

   The gravitational wave detector network, including GEO600 in Germany, collects a TByte of data each day, which must be searched using different algorithms for possible events such as black hole or neutron star collisions, or pulsar signals.

   Routine realtime analysis of gravitational wave data from the Hanover detector identifies a burst event, but this standard analysis reveals no information about the burst location. To obtain the location, desperately required by astrophysicists for turning their telescopes to view the event before it fades, a large series of templates must be cross-correlated against the detector data. An Italian astrophysicist accesses the GEO600, and using the performance tool finds that 3 TFlops/s is needed to analyze the 100GB of raw data in the required hour. Local resources are insufficient, so using the brokering tool, she locates the fastest available machines around the world. She selects five suitable machines, and with scheduling and data management tools, data is moved, executables created and the analysis starts. In an Amsterdam bar twenty minutes later, an SMS message from the portal's notification tool, informs her that one machine is overloaded, breaking the runtime contract. She connects with her PDA to the portal, and instructs the migration tool to move this part of the analysis to a different machine. Within the specified hour, a second SMS message tells her analysis is finished, and the resulting data is now on her local machine. Using this location data, observatories are able to find and view an exceptionally strong gamma-ray burst, characteristic of a collision of neutron stars.

2. Numerical Relativity:

   A single simulation of an astrophysical event, e.g. black hole or neutron star collisions, ideally requires over TByte and TFlop resources, not yet available on a single machine.

   Learning more about the detected burst requires cross-correlating detector data with custom wave templates from full-scale neutron star simulations. Sufficiently accurate templates require running large scale simulations too big to fit on any current supercomputer. German members of an international numerical relativity collaboration are tasked with creating collision templates for ten different neutron star mass combinations. They access the web-based Simulation Portal, selecting required code modules and building parameter files with the code composition tool. The performance prediction tool estimates that each simulation requires 1024 GigaBytes of memory and 100,000 GigaFlops, with an additional 500,000 GigaFlops required for processing data to create signal templates. The brokering tool finds that no single machine in the Simulation Testbed can supply enough memory, but locates two machines which can be connected to form a large enough virtual supercomputer, the dynamic grid monitoring tool indicates an acceptable bandwidth between them. The scheduling tool stages the five runs to appropriate queues on the machines, and the first simulation starts. The time-consuming task of creating templates is handled by spawning simulations to smaller machines dynamically located by the broker, at each time step data is streamed to a series of networked computers for analysis, creating a simulation vector using available machines on the grid. Collaborators around the world connect to the portal using networked workstations, home PCs and modems, as well as the latest wireless PDAs and mobile phones. They are able to use various remote access tools to visualize data, monitor performance and simulation properties, and interactively steer the simulation.