CloudMirror: Application-Aware Bandwidth Reservations in the Cloud

Cloud computing providers today do not offer guarantees for the network bandwidth available in the cloud, preventing tenants from running their applications predictably. To provide guarantees, several recent research proposals offer tenants a virtual cluster abstraction, emulating physical networks. Whereas offering dedicated virtual network abstractions is a significant step in the right direction, in this paper we argue that the abstractions exposed to tenants should aim to model tenant application structures rather than aiming to mimic physical network topologies. The fundamental problem in providing users with dedicated network abstractions is that the communication patterns of real applications do not typically resemble the rigid physical network topologies. Thus, the virtual network abstractions often poorly represent the actual communication patterns, resulting in overprovisioned/wasted network resources and underutilized computational resources. We propose a new abstraction for specifying bandwidth guarantees, which is easy to use because it closely follows application models; our abstraction specifies guarantees as a graph between application components. We then propose an algorithm to efficiently deploy this abstraction on physical clusters. Through simulations, we show that our approach is significantly more efficient than prior work for offering bandwidth guarantees.