Large-scale data processing software and performance instabilities within HEP grid environments

Large tasks running on grids and clouds have introduced a need for stability guarantees from geographically spanning resources, where failures are handled pre-emptively. Detecting performance inefficiencies in such cases is difficult. While individual services implement fault-tolerance, the behaviour of interacting failures within tightly-coupled systems is less understood. This paper describes an approach to modelling performance of production tasks running within the ALICE grid. We provide an overview of the ALICE data and software workflow for production jobs. Event states are then constructed, based on data centre job, computing, storage and user behaviour. We then address the question of analysing failures within the context of operational instabilities, occurring in production grid environments. The results demonstrate that operational issues can be detected and described according to the principle service layers involved. This can guide users, central and data centre experts to take action in advance of service failure effects.