Abstract:  

To improve user-perceived performance, large Internet content providers
(e.g., Google, Microsoft, Facebook) are expanding their private network
clouds towards end users. In particular, they are aggressively pursuing
direct peering with last-mile consumer networks, thus making their
content closer to end-users. In addition, such clouds host a number of
applications that enable clients to actively upload user-generated
content to the cloud. Moreover, they agilely replicate such content to
one or multiple points in the cloud, thus making it closer to
anticipated recipients. In this work, we explore techniques for
utilizing these rapidly-growing clouds, their openness to content
uploads and their internal data replication mechanisms for the purpose
of designing a generic data transfer architecture on top of such clouds.
Our contributions are threefold. First, we conduct a broad measurement
study of the most popular content providers and applications including
Youtube, Flickr, Hotmail, and Gmail, and effectively characterize the
underlying cloud infrastructures as well as their replication
mechanisms. Second, we design and implement Skynet, a data transfer
network architecture that optimizes end-user performance by
opportunistically selecting the underlying clouds to transfer data over.
Finally, we extensively evaluate Skynet and demonstrate that both direct
and cloud-hopping Skynet paths are capable of achieving dramatic
improvements relative to regular Internet paths. Such cloud paths are
stable over longer time scales. Hence, they can become attractive
alternatives for numerous endpoint applications.