1 Resource Fault Prediction Models

Abstract. The use of tools to forecast faults over computational resources composing highly distributed system, like Grid systems, represents an interesting approach to control the comprehensive performance of these systems, limiting the number of job submissions to those computational resources temporally unable to complete any job. The analysis of log information retrieved from a real system shows an homogeneous behavior of the computational resources involved, and therefore the possibility to build forecasting tools limiting the number of job submissions that fail. The results obtained in this paper show on one side the possibility to obtain high percentage of correct forecasts even with simple models, and from the other side the difficulty to improve these results even using adaptive models relatively complex