An Architecture for an Adaptive Run-time Prediction System

This article describes a system for run-time prediction of applications in heterogeneous environments. To exploit the power of computational grids, scheduling systems need profound information about the job to be executed. The run-time of a job is - beside others - not only dependent of its kind and complexity but also of the adequacy and load of the remote host where it will be executed. Accounting and billing are additional aspects that have to be considered when creating a schedule. Currently predictions are achieved by using descriptive models of the applications or by applying statistical methods to former jobs mostly neglecting the behaviour of users. Motivated by this, we propose a method that is not only based on the characteristics of a job but also takes the behaviour of single users and groups of similar users respectively into account. The basic idea of our approach is to cluster users, hosts and jobs and apply multiple methods in order to detect similarities and create forecasts. This is achieved by tagging jobs with attributes and by deriving predictions for similar attributed jobs whereas the recent behaviour of a user determines which predictions are finally taken.