Speaker
Description
Impact
The ATLAS detector is going to take real data in Autumn 2009. Spanish sites have to be validated in order to be ready for the physics analysis. With these tests we will identify breaking points and bottlenecks which result from the site design or configuration taking into account that needs of analysis jobs differ from those of production. The essential impact is ensuring the robustness and effectiveness of the complex system of GRID Analysis
Conclusions and Future Work
Distributed analysis tests are necessary to stress the facilities at a simulated full user load. Very useful first test that allows to identify “big” problems (missing software, missing/bad information published) and already to point out some limitation (e.g.:network connection in some sites). In the future, we would like to test new parameters to improve access to the data using DCAP and LUSTRE and reproduce the exercise with more jobs to really hammer the sites in order to stress the system.
Keywords
Distributed Analysis, GRID, LCG/EGEE; ATLAS, GANGA, LUSTRE, DCAP
Detailed analysis
The needs of analysis jobs differ from those of production, so though a site may function for MC, it may have a poor performance for analysis. With these tests we are trying to identify what works well, and what doesn’t work well. The distributed analysis challenge framework is built with Ganga LCG/EGEE jobs, which submits an application to the sites, tracks the job statuses periodically as they run, and allows us to easily develop reports after the complete. The jobs read directly from the Storage Element the data using posix I/O and some performance metrics, for instance success/failure rate, CPU/walltime, events per second and error classification, e. g.: different I/O errors are recorded. The results led to discovery network and switch problems and software bugs. The same application and input data were used in all sites.
URL for further information
http://gangarobot.cern.ch/st/
