Une approche basée sur les machines volontaires pour le calcul haute performance

Abstract

High Performance Computing (HPC) systems aim to solve complex computing problems (in a short time) that are either too large for standard computers or would take too much time. Solutions based on supercomputers, clusters, grids and cloud are not suitable for everyone because their costs generally exceed funding (particularly for academics) and the administrative complexity of access creates a higher barrier. In this work, we propose to use idle computing resources of volunteers to build very low-cost HPC systems suitable for resource-poor settings found mainly in developing countries such as Cameroon. This approach is based on a hybrid architecture in which the server acts as a global information directory and a set of computers acting as volunteers and clients. This approach requires no additional hardware investment, but builds on devices already owned by users and their communities. It therefore guarantees autonomy and practically free access. It should be noted, however, that computing resources are generally shared with owners and with power and internet outages, the availability of volunteer computers cannot be guaranteed for any period of time. A volunteer on which a task is in running may become unavailable at any time without prior warning and, consequently, might have several negative effects, including loss of data or computation. Fault tolerance techniques, such as redundancy or checkpointing, increase significantly deployment cost. To reduce these losses, we use the idea of selecting the resources most likely to be available until the end of each task. Machine learning methods are used to build resource availability prediction models. The evaluation of three predictors (naive Bayes classifier, Lasso regression and multilayer perceptron) allowed us to observe that none of them is better than the others in terms of average prediction accuracy. To maintain the benefits of each predictor, we provide a resource availability prediction system based on model selection. We apply it for a classification method (the multilayer perceptron) and a linear regression method (the Lasso method). In addition to volunteer computing systems, we used the resource availability prediction model to improve the performance of recommendation systems. Finally, we have successfully implemented an HPC system called VC_UY1 based on the personal computers of students and lecturers at the University of Yaoundé I. This system consists of a server and 164 volunteer computers. It provides a theoretical power of 12 PFlops. We have conducted experiments on the mathematical problem of the multiplication of large matrices and on the physical problem of the characterization of stochastic resonance. The results show that volunteer computing systems can provide significant computing power at a very low cost. The resource availability prediction concept has been extended to recommendation systems to determine which products are likely to be purchased at a given time. Experiments conducted on three recommendation system datasets (MovieLens-2k, MovieLens-ls and Last.fm) have shown that product purchase prediction improves recommendation results.