Política de escalonamento de recursos computacionais em clusters de física de altas energias

Abstract

This paper presents a proposal to allow the shared use of computing resources available in a cluster in order to support simultaneously the four experiments at CERN. The approach uses the concept of contracts, where the requirements and restrictions of each experiment are described in profiles, and a resource allocation policy is set to maintain the use of resources to meet those profiles. We propose an architecture model to manage the shared use of a cluster by the four LHC Virtual Organizations. This architectural model is composed of elements common to a typical Tier-2 cluster plus features to control the admission of new jobs from all LHC Virtual Organizations. This model monitores the resource usage for each cluster VO, allocates resources to each job according to an established policy for the cluster in order to seek compliance with requirements of each VO. We define an algorithm for scheduling jobs, that uses preemption mechanisms for controlling the allocation of cluster nodes depending on the current use of resources and accumulated by each VO. This algorithm runs in one of the elements of architecture named broker, since it mediates the influx of new jobs and the allocation of the cluster nodes, and is aiming to keep control of contracts for each VO. The simulation of the proposed architecture was done in GridSim simulator grid and the results were evaluated.