Osprey: Operating system for predictable clouds

Cloud computing is currently based on hardware virtualization wherein a host operating system provides a virtual machine interface nearly identical to that of physical hardware to guest operating systems. Full transparency allows backward compatibility with legacy software but introduces unpredictability at the guest operating system (OS) level. The time perceived by the guest OS is non-linear. As a consequence, it is difficult to run real-time or latency-sensitive applications in the cloud. In this paper we describe an alternative approach to cloud computing where we run all user applications on top of a single cloud operating system called Osprey. Osprey allows dependable, predictable, and real-time computing by consistently managing all system resources and exporting relevant information to the applications. Osprey ensures compatibility with legacy software through OS emulation provided by libraries and by porting runtime environments. Osprey's resource containers fully specify constraints between applications to enforce full application isolation for real-time execution guarantees. Osprey pushes much of the state out of the kernel into user applications for several benefits: full application accounting, mobility support, and efficient networking. Using a kernel-based packet filter, Osprey dispatches incoming packets to the user application as quickly as possible, eliminating the kernel from the critical path. A real-time scheduler then decides on the priority and order in which applications process their incoming packets while maintaining the limits set forth in the resource container. We have implemented a mostly complete Osprey prototype for the x86 architecture and we plan to port it to ARM and PowerPC and to develop a Linux library OS.