For the fifth year, experimenters and members of Scientific Computing Division (SCD) gathered for the annual FIFE Workshop. The workshop focus was divided between discussions of the FIFE roadmap on the first day and extensive tutorials on the second day. The workshop had more than 65 attendees from across all Frontiers (Intensity, Cosmic, and Energy) and from all departments within SCD. The agenda and contributions to the workshop can be found on the FIFE Workshop Indico page. While the series had very productive discussions, the FIFE group has decided that this will be the last summer in the series for the foreseeable future.
The most interesting aspect was the Roadmap discussion concerning the future of FIFE services and the direction of experiment computing model development. The hottest topic of the day was the incorporation and availability of GPU hardware into Open Science Grid resources. With Deep Learning becoming an essential part of offline and online computing for experiments, the availability of large scale GPU clusters has become an issue for researchers developing new and novel Deep Learning algorithms. Other topics of discussion included the refactoring of FermiGrid (previously known as GPGrid) in order to take full advantage of node resources and establish needed infrastructure changes to allow for the use of Docker images and future use of Singularity on FermiGrid. In addition, plans for modifications to the Data Storage infrastructure and unmounting of BlueArc volumes (e.g., /<experiment>/app) from worker nodes were discussed. The final goal of the FIFE Roadmap is the update of services in order to mesh with the dynamic resource allocation of HEPCloud that is planned to go live at the end of 2018. The Roadmap discussion was successful enough that the FIFE Group plans to hold biannual roadmap discussions with Offline and Production Coordinators in December and June going forward.
The second day of the workshop focused on tutorials for FIFE services and best practices discussions so that users could get the greatest utilization of computing resources and produce physics results as quickly as possible. Tutorial topics included SAM4Users, job submission, workload management with POMS, data handling, and updated monitoring through FIFEMon. These important tutorials are still available from the FIFE workshop Indico page, and can be used as the basis for experiment specific tutorials on FIFE services. One of the best attended tutorials was the updated SAM4Users tutorial which gives an overview of the tools available for individual users to register files in SAM, create datasets, and archive to tape storage without the heavy overhead of creating complete file metadata. Also, best practice talks discussing profiling workflows, optimizing job submission, and troubleshooting failures were presented. Since we have established a good working set of tutorials, experiments can request that the FIFE group present these tutorials for their collaborators at any time. We look forward to working with experiments to help get data analyzed and published as soon as possible.