An AI-driven Resource-Recommendation System for Cloud HPC

Abstract

Ιn recent years, the Cloud has evolved into a powerful environment with many instances available for the user to choose. Especially, for High-Performance Computing (HPC) applications the vast amount of available hardware leaves the user with a challenging choice to make. When making this choice, many aspects should be taken into account, in particular the available budget and the time-frame of the project. This information, although fundamental, is really hard to acquire before project completion. In this thesis, a Resource-Recommendation System, based on Neural-Networks (NNs), is proposed in order to help the user determine which instance fits their needs best and help them make their choice. For these reasons, different Multi-Layer Perceptrons (MPLs) where selected and evaluated on different High-Performance Computing (HPC) benchmarks, both synthetic and application. Initially, raw-data was collected from different Cloud instances for each benchmark and organised in a proper way to later be used. The next phase, was the data-prepossessing, in which each data-set was surveyed in order to determine the available trends and a synthetically-generated, more complete data-set was created for the majority of the benchmarks. Lastly, the new syntheticallygenerated data-set was encoded in a proper way and it was fed into the Neural Network (NN) in order to predict the execution times of each benchmark. In total, a combination of six different benchmarks and five Multi-Layer Perceptros (MLPs) were trained and evaluated. Last but not least, the training of the Neural-Networks (NNs) was realised both on Graphics Processing Units (GPUs) and Graphcore’s Intelligence Processing Units (IPUs).