Towards a Workload Mapping Model for Tuning Backing Services in Cloud Systems.

With the increasing advent of applications and services adopting cloud-based technologies, generic automated tuning techniques of database services are gaining much attraction. This work identifies and proposes to overcome the potential challenges associated with deploying a tuning service as part of Platform-as-a-Service (PaaS) offerings for tuning of backing services. Offering an effective database tuning service requires such tuners whose architecture can support tuning multiple databases and numerous database versions deployed on various types of underlying hardware configurations with varying VM plans. Tuners that offer such capabilities usually attempt to leverage experiences gathered previously. By taking advantage of relevant past experiences, tuners classify the current workload to the most pertinent workload seen recently. In this work, a five-layered, fully connected neural network with ReLU activation function is being employed as the classification model to classify data points into relevant workload classes. The categorical cross-entropy function is employed as the loss function and optimized using Adam optimizer. The work handles the challenges related to the cold-start problem, issues in mapping, and cascading errors. The proposed solution can overcome these issues in a large-scale production environment. The results show that the model has 93.3% accuracy in 93.8% F1-score as compared to the previous model like Ottertune.