Predictive modeling for management of database resources in the cloud

Abstract

Public cloud providers who support a Database-as-a-Service offering must efficiently allocate computing resources to each of their customers in order to reduce the total number of servers needed without incurring SLA violations. For example, Microsoft serves more than one million database customers on its Azure SQL Database platform. In order to avoid unnecessary expense and stay competitive in the cloud market, Microsoft must pack database tenants onto servers as efficiently as possible. This thesis examines a dataset which contains anonymized customer resource usage statistics from Microsoft's Azure SQL Database service over a three-month period in late 2014. Using this data, this thesis contributes several new algorithms to efficiently pack database tenants onto servers by collocating tenants with compatible usage patterns. An experimental evaluation shows that the placement algorithms, specifically the Scalar Static algorithm and the Dynamic algorithm, are able to pack databases onto half of the machines used in production while incurring fewer SLA violations. The evaluation also shows that with two different cost models these algorithms can save 80% of operational costs compared to the algorithms used in production in late 2014.