Speaker
Gerardo Ganis
(CERN/SFT)
Description
The goal of PROOF (Parallel ROOt Facility) is to enable the interactive analysis of distributed data sets in a transparent way. PROOF represents an alternative to a batch-oriented usage of a distributed computing system. The system is being successfully used by the PHOBOS experiment since a few years.
In view of the startup of the LHC analysis PROOF underwent several developments and improvements mainly with the purpose to adapt it to the large multi-user analysis environments found at LHC. The ALICE collaboration has pioneered test-stressing of the system using the first CAF (Central Analysis Facility), composed by 40 machines located at CERN. The test-CAF represents a realistic environment for LHC analysis, with a non-negligible amount of local disk space managed by the XROOTD daemon populated with dedicated upload facilities (e.g. on-demand stage-in from the CASTOR system).
The ALICE tests have brought in several new developments, among which:
* augmentation and fine-tuning of package management and environment setting facilities;
* a new "packetizer" (the system element responsible for work distribution) optimizing the load balancing algorithm at the level of a single query; the new packetizer minimizes network transfers, reducing hot spots in accessing a specific data server and avoiding in this way bottleneck effects at the end of jobs. A new dynamic way of calculating packet sizes as a time quantum has also been introduced.
* the integration of the system with a user-scheduling facility with the main purpose of increasing fairness in the cluster usage and to provide a handle for the implementation of quota policies.
In this talk we will describe in detail the status of the new developments and outline the future plans to bring PROOF in production for LHC analysis.
