First stage of the ATLAS Data Challenge on the NorduGrid
In July-August 2002, the NorduGrid testbed and middleware were successfuly tested in the production environment, being the main tool in the Nordic participation in the ATLAS Data Challenge 1, phase 1.
The purpose of the ATLAS Data Challenges (DC) is to prepare for the data taking and analysis by the ATLAS detector at the future LHC accelerator at the European Organization for Nuclear Research (CERN). The first phase of the ATLAS DC consists of applying detector simulation algorithms onto computer-generated events of proton-proton collision. This task was distributed between all the members of the ATLAS collaboration, and the Nordic share comprised 15 out of 150 partitions of the so-called "dataset 002000" (di-jet events).
Particle physicists in Nordic countries are commited to use Grid tools in their activities, being the first particle physics community to use Grid in the daily work via the NorduGrid testbed. Therefore, it was natural to execute the ATLAS Data Challenge assignment by making use of the NorduGrid.
The NorduGrid toolkit is a Globus-based Grid middleware, developed to meet the basic requirements of distributed computing and data handling. It was installed and tested on a variety of Linux clusters, and can be downloaded from the NorduGrid site at www.nordugrid.org. Its most important components are:
- the MDS-based Information System, providing reliable dynamic information about the status of the services
- the Grid Manager, residing at a gatekeeper and performing job pre- and post-processing, as well as job submission to the local resources
- the User Interface, responsible for the user's interaction with the Grid, including resource brokering, communication with the Grid Manager and job monitoring
The architecture allows any kind of jobs to be submitted and optimally distributed over the NorduGrid testbed. It also has efficient stage-in and stage-out procedures, with possibility to use most data transfer protocols, such as GridFTP, FTP and HTTP.
 |
The ATLAS DC1/1 jobs on the NorduGrid were submitted in batches via the User Interface, which peformed job requirements matching, located optimal resources and sent requests to chosen clusters. The Grid Manager downloaded all the necessary input and automatically uploaded the output to the Storage Element in Oslo, registering the metadata to the Replica Catalog. Job monitoring was done by using either the Web or a command-line interface to the Information System.
The technical challenge was to process ca. 30GB of input data in 300 partitions, each producing about 1GB of output. Each input partition contained around 5000 events, of which ca 400 had to be fully processed. Average time to process a single event on the NoirduGrid testbed was 150 seconds, which means that the total CPU time required to complete the exercise was 2×107 seconds. Total number of computing nodes participating in the Nordic ATLAS Data Challenge was 49, located at 5 clusters: in Denmark (Copenhagen), Norway (Bergen and Oslo) and Sweden (Lund and Uppsala). Produced output exceeded 300GB, 20% of which was uploaded to CERN for further processing in the coming Data Challenges.
The challenge was successfully met, with the failure rate being negligibly small, with all the failures being attributed not to the NorduGrid middleware, but to a third-party software. A summary of the performed task can be found at the dedicated page of the NorduGrid Web-site.
In future, it is foreseen to continue running ATLAS DataChallenges on the NorduGrid, using upgraded middleware and increased computing resources, e.g., new clusters in Sweden (Stockholm and Linköping).