Predição de tempo de execução de tarefas em grades computacionais para recursos não dedicados

Abstract

Runtime prediction of jobs to grid resource is an important data to schedulers and brokers. Grids configure environment composed by distinct and several resources and users, with different kind of jobs, do not have a pattern to submission of job, and hence runtime prediction of job is a large challenge. This work explore the problem of runtime prediction of jobs to non-dedicated resources. We treat this problem based in a methodology proposal in Ph.D. thesis of Lilian Noronha Nassif, named PredCase, it showed a efficient methodology for dedicated resources. In our method, named NdrPredCase, we utilize past cases of runtime jobs to calculate the runtime prediction to a new job, how done in PredCase. We utilized the paradigm Case-Based Reasoning to develop the NdrPredCase phases: retrieval cases alike to new job, reuse this cases to calculate the initial solution, adaptation of solution to a forecast workload and retain information about job description, solution description and running data. The main contribution of this work is the adaptation of a initial solution, it calculated with past cases, using a relation of previous workload and forecast workload to running job. Results of experiments demonstrated that NdrPredCase has a good accuracy to calculate runtime prediction to non-dedicated resources if the amount of past cases is sufficient, and that NdrPredCase has a good performance to calculate this prediction.