We consider a Netflix-like video-on-demand (VoD) system, where the video popularity remains stable over a day or week while the request traffic may vary significantly within a day (e.g., by orders of magnitude). To respond to dynamic user traffic in a timely manner, we study an auto-scaling cloud-based VoD data center, where the virtual machines (VMs) can be turned on and off according to user traffic to elastically scale system resources. Movies are stored in persistent storage as standby unit, which is attached to VMs on-demand at any time. Due to limited VM streaming capacities and persistent storage size, a movie request is dispatched either to an operating VM or to a remote repository (the so-called remote traffic). We are interested in minimizing the number of operating VMs given a...[ Read more ]

We consider a Netflix-like video-on-demand (VoD) system, where the video popularity remains stable over a day or week while the request traffic may vary significantly within a day (e.g., by orders of magnitude). To respond to dynamic user traffic in a timely manner, we study an auto-scaling cloud-based VoD data center, where the virtual machines (VMs) can be turned on and off according to user traffic to elastically scale system resources. Movies are stored in persistent storage as standby unit, which is attached to VMs on-demand at any time. Due to limited VM streaming capacities and persistent storage size, a movie request is dispatched either to an operating VM or to a remote repository (the so-called remote traffic). We are interested in minimizing the number of operating VMs given a certain remote traffic constraint, by jointly optimizing which movies to store in each persistent storage unit, which storage units to be attached for online VMs, and which VMs to dispatch user requests.

We formulate the problem and show that it is NP-hard. We propose AMATO (Auto-scaling Movie Allocation and Traffic Optimization), an efficient approximation algorithm which achieves 2-approximation at peak traffic. Both experiments on cloud platform (Amazon EC2 ) and trace-driven simulation based on large-scale real-world data show that AMATO is closely optimal. It achieves significantly lower number of operating VMs as compared with the other state-of-the-art and traditional schemes.