No. 115 (00239) Family name : Garzoglio Given name : Gabriele Affiliation : Fermilab Abbreviation : E-mail address : garzoglio@fnal.gov Title : A local sandbox management for distributed applications Authors : Gabriele Garzoglio, Igor Terekhov, Andrew Baranovski, Lee Lueking, Parag Mhashilkar, Vijay Murthi Abstract : The SAM-Grid project provides a job management and data handling infrastructure that enables globally distributed computing for the DZero and CDF experiments at Fermilab. One of the major challenges for the deployment of our system, is coping with the diversity of the local job submission interfaces. In particular, once a job has been dispatched to the gateway of a computing cluster, the transport of the input/output sandbox and data to/from the worker nodes is implemented locally with a variety of approaches. Some local Batch Systems are configured to manage stage-in/stage-out of files, some rely on a shared file system for the head-worker nodes communication, some rely on having the worker nodes on an open network. Depending on the hardware configuration, it may be most practical either unpacking the input sandbox at the head node, or storing it compressed there and unpacking it at the worker nodes. Some of these resources are dedicated to the experiments, some are completely general purpose. Furthermore, the necessity of transporting Gigabytes of input data, generally not present in the proximity of the cluster at the time of submission, raises the level of complexity. The sandbox management infrastructure of the SAM-Grid provides a level of abstraction for the possible local configurations and implements techniques (e.g. dynamically instanciated file transport daemons such as gridftpd) to implement a configurable and reliable local input/output data management. In this paper we discuss such abstractions and a categorization of the local configurations after our experience in the deployment of the SAM-Grid to a dozen different sites.