Time series forecasting of application resource usage applying deep learning methods

Improving the efficiency of big cloud providers has become a very difficult task. The great quantity of workloads that are run on the cluster nodes and their wide diversity and heterogeneity extensively complicates it. One of the main issues is the divergence between the requested and the real usage of the workloads. This difference causes the nodes to not efficiently use all of their computing resources. Past works have tried to tackle this problem forecasting the future usage of the workloads and dynamically changing their allocated resources or creating/removing new replicas. However, they have failed to properly predict the correct resource usage during the high intense moments of consumption, i.e., the spikes of usage. These prediction errors can cause heavy problems of resource starvation in the cluster nodes and heavily diminish the quality of service of the cloud provider. Also, the majority of contributions use metrics that are not suited for the specific case of cloud provisioning, that do not not quantify properly the prediction error during the spikes of usage of the workloads. For this reason, in this work, I am proposing mainly two new contributions in this regard. Firstly, a new approach to forecast the future resource consumption of workloads with the help of deep learning models that has demonstrated a good performance in the specific situations of high intensive moments of resource usage. Secondly, a new evaluation that has proven to correctly quantify the quality of the predictions in traces that contain a notable number of spikes. The prior contributions can help in improving the scheduling on the cluster nodes and the good management in the task of sharing resources between multiple workloads, improving the final resource efficiency of the cloud provider