In an auto-scaling cloud-based video-on-demand (VoD) data center, the virtual machines (VMs) can be turned on and off dynamically to serve fluctuating user traffic more cost-effectively. In such system, user traffic is mapped to one of the "auto-scaling levels" where a certain subset of VMs are turned on. These VMs are pre-loaded with movies to save the cost of movie replication at level switch. Due to limited storage and streaming capacities, the VMs may not serve all the movie requests; in which case the requests are directed to a repository as misses.

We study, for the first time, the joint optimization of allocating movies and distributing user traffic among the VMs so as to minimize the misses at the repository. We formulate the problem and prove that it is NP-hard. We then...[ Read more ]

In an auto-scaling cloud-based video-on-demand (VoD) data center, the virtual machines (VMs) can be turned on and off dynamically to serve fluctuating user traffic more cost-effectively. In such system, user traffic is mapped to one of the "auto-scaling levels" where a certain subset of VMs are turned on. These VMs are pre-loaded with movies to save the cost of movie replication at level switch. Due to limited storage and streaming capacities, the VMs may not serve all the movie requests; in which case the requests are directed to a repository as misses.

We study, for the first time, the joint optimization of allocating movies and distributing user traffic among the VMs so as to minimize the misses at the repository. We formulate the problem and prove that it is NP-hard. We then propose AMTO (Auto-scaling Movie allocation and Traffic Distribution Optimization), a novel and efficient algorithm to address the problem. AMTO achieves computational efficiency even for large movie pool. Extensive simulation shows that it is closely optimal, achieving significantly lower miss traffic as compared with state-of-the-art schemes (often by more than 50%).