A Three-Phase Adaptive Prediction System of the Run-Time of Jobs Based on User Behaviour

摘要

This article describes an approach for predicting the run-time of jobs using a technique that works in three phases. Each one is independently adjusting to a user's behaviour in order to lead to accurate forecasts. In heterogeneous and distributed environments it is necessary to create schedules for utilizing the resources in an efficient way, but the generation of these schedules often poses a problem for a scheduler, as it has to incorporate several aspects like priorities, system load, Service Level Agreements. One possibility to support a scheduler in doing its work is to provide accurate predictions of the run-times of the submitted jobs.A large number of current techniques for run-time prediction offer statistical models - in the majority of cases linear ones - that are deployed on previously filtered data. As users have different jobs due to their field of work, and the attributes of their jobs differ, because of the different requirements they have, filtering data and choosing an appropriate method for a forecast has to cover these aspects. Motivated by this we propose an adaptive prediction system, where in each one of the phases we adjust our methodology on basis of the former behaviour of a user. This leads to a user specific clustering of data and to a flexible utilization of different prediction techniques in order to create a user-centred prediction model.