Wen Guan

Distributed analysis with PROOF in ATLAS collaboration

The Parallel ROOT Facility – PROOF is a distributed analysis system which allows to exploit inherent event level parallelism of high energy physics data. PROOF can be configured to work with centralized storage systems, but it is especially effective together with distributed local storage systems – like Xrootd, when data are distributed over computing nodes. It works efficiently on different types of hardware and scales well from a multi-core laptop to large computing farms. From that point of view it is well suited for both large central analysis facilities and Tier 3 type analysis farms. PROOF can be used in interactive or batch like regimes. The interactive regime allows the user to work with typically distributed data from the ROOT command prompt and get a real time feedback on analysis progress and intermediate results. We will discuss our experience with PROOF in the context of ATLAS Collaboration distributed analysis. In particular we will discuss PROOF performance in various analysis scenarios and in multi-user, multi-session environments. We will also describe PROOF integration with the ATLAS distributed data management system and prospects of running PROOF on geographically distributed analysis farms.

Data-oriented scheduling for PROOF

The Parallel ROOT Facility - PROOF - is a distributed analysis system optimized for I/O intensive analysis tasks of HEP data. With LHC entering the analysis phase, PROOF has become a natural ingredient for computing farms at Tier3 level. These analysis facilities will typically be used by a few tenths of users, and can also be federated into a sort of analysis cloud corresponding to the Virtual Organization of the experiment. Proper scheduling is required to guarantee fair resource usage, to enforce priority policies and to optimize the throughput. In this paper we discuss an advanced priority system that we are developing for PROOF. The system has been designed to automatically adapt to unknown length of the tasks, to take into account the data location and availability (including distribution across geographically separated sites), and the {group, user} default priorities. In this system, every element - user, group, dataset, job slot and storage - gets its priority and those priorities are dynamically linked with each other. In order to tune the interplay between the various components, we have designed and started implementing a simulation application that can model various type and size of PROOF clusters. In this application a monitoring package records all the changes of them so that we can easily understand and tune the performance. We will discuss the status of our simulation and show examples of the results we are expecting from it.