Andrew D. Gordon

"Service Combinators for Farming Virtual Machines"

(Vortrag im Rahmen der "Distinguished Lecture Series 07/08" des Max Planck Instituts für Software-Systeme)

We consider the problem of managing server farms largely automatically, in software. Automated management is gaining in importance with the widespread adoption of virtualization technologies, which allow multiple virtual machines per physical host. We consider the case where each server is service-oriented, in the sense that the services it provides, and the external services it depends upon, are explicitly described in metadata. We describe the design, implementation, and formal semantics of a library of combinators whose types record and respect server metadata. Our implementation consists of a typed functional script built with our combinators, in control of a Virtual Machine Monitor hosting a set of virtual machines. Our combinators support a range of operations including creation of virtual machines, their interconnection using typed endpoints, and the creation of intermediaries for tasks such as load balancing. Our combinators also allow provision and reconfiguration in response to events such as machine failures or spikes in demand. We describe a series of programming examples run on our implementation, based on existing server code for order processing, a typical data centre workload. To obtain a formal semantics for any script using our combinators, we provide an alternative implementation of our interface using a small concurrency library. Hence, the behaviour of the script plus our libraries can be interpreted within a statically typed process calculus. Assuming that server metadata is correct, a benefit of typing is that various server configuration errors are detected statically, rather than sometime during the execution of the script.