Using the EMMSP model to predict the available computing resources in the cluster systems

Selecting the environment for computation can be quite challenging if you want to estimate which of the available computing environments will complete computations earlier. The solution of this and several related tasks has provoked many studies aimed at optimizing the use of computing resources. This paper describes the EMMSP prediction model, which provides an effective solution for the problem of predicting the available computing resources. Previously, this model has worked well for the problem of electricity price forecasting. The purpose of this paper is to evaluate the strengths and weaknesses of the model in relation to the problem of forecasting the available computing resources, as well as the possibility of its combination with other models. The application analysis focuses on the interpretability of results and the possibility of taking management decisions on their basis. The EMMSP forecasting model is integrated with the Templet Web analytics cloud service subsystem, designed for automation of scientific computing on the basis of the Sergey Korolev cluster at the Samara University. The analytics subsystem performs continuous forecasting of the cluster load and provides prediction data in tabular and graphical form for service users. The main task of the subsystem is to forecast the number of available resources of different types for 12 hours ahead, one forecast point per hour. In addition to direct prediction, the analytics subsystem searches and analyzes patterns in the time series of the cluster load for retrospective analysis. The evaluation the applicability of the model and forecasting errors was performed using the load statistics from the Sergey Korolev cluster collected in the period from November 2013 to May 2016. To show the possibility of combination EMMSP with other prediction models the paper displays the improvement of the EMMSP model prediction results when used in adaptive combination with a naive forecasting model based on the time series data shift. The resulting adaptive model gives a smaller number of prediction errors than its separate components. The results of forecasting available resources using the EMMSP model can be used to solve various problems, such as designing the placement of distributed application components, optimization of the startup parameters and volumes of input data, the reduction of energy consumption and the planning the maintenance periods of the cluster nodes.