Predictive Provisioning: Efficiently Anticipating Usage in Azure SQL Database

Seagull: An Infrastructure for Load Prediction and Optimized Resource Allocation

“…Load Prediction for optimized resource allocation on a cluster has become a popular research direction in the recent years. Existing approaches focus on predicting survivability of databases for optimized resource provisioning [34], idle time detection for database quiescing and overbooking [27,39], database workload prediction for database consolidation [18], VM workload prediction [25] for oversubscribing servers [14], dynamic VM provisioning [13], and reducing performance interference between VMs co-located on the same physical machine [32], workload classification for capacity planning and task scheduling [31], cost-and QoS-aware application placement in virtualized server clusters [38,40], and preemptive auto-scale of resources [19,20,21,22,26,36,33,35,37,41]. None of these approaches focused on predicting low load windows for optimized scheduling of system maintenance tasks.…”

“…Microsoft Azure, Google Cloud Platform, Amazon Web Services, and Rackspace Cloud Servers are the leading cloud service providers that aim to guarantee high quality of service to their customers, while controlling operating costs [28,39]. Achieving these conflicting goals manually is laborintensive, time-consuming, error-prone, neither scalable, nor durable.…”

“…While time series forecast in general and load prediction in particular are well studied topics, none of the state-of-the-art approaches focused on predicting the lowest valley in customer CPU load for optimized backup scheduling. Instead, existing approaches focus on, for example, idle time detection for predictive resource provisioning [27,39], VM workload prediction for dynamic VM allocation [13,14], and demand-driven auto-scale of resources [19,20,21,22,36,37,41]. Thus, these approaches do not tackle the unique challenges of low load prediction for optimized backup scheduling described above.…”