Virtual Machine Mobility With Self-Migration

Preface This work is a continuation of the NomadBIOS project, where Asger Jensen and I designed and built what would now be termed a para-virtualized hypervisor for hosting multiple virtual machines on top of the L4 microkernel, and—to our knowledge—were the first to show the feasibility of live migration of virtual machines across a local network. Focusing on grid computing, our idea was to package each grid job in a VM, and have these VMs roam the grid in search of idle resources. This work is described in our joint Master's thesis " Nomadic Operating Systems " [Hansen and Henriksen, 2002]. When starting my Ph.D.-studies about a year later, a group at the University of Cam-bridge had just released the first version of their own hypervisor, called Xen. Compared to the L4 microkernel on which our work was based, the Xen VMM had a number of advantages; it was simpler and faster, came with built-in drivers for up-to-date hardware, and because it was explicitly focused on hosting virtual machines, required less effort when porting or maintaining guest operating systems. Only a few weeks into my stipend, I travelled to Cambridge and started cooperating with the Xen team on bringing the live-migration features developed for NomadBIOS to Xen. While struggling to understand the design of their hypervisor, it occured to me that live VM migration would require a redesign of Xen to make it more like L4, compromising some of Xen's simplicity. Inspired by Steven Hand's " Self-paging in Nemesis " [Hand, 1999] paper, it occured to me that if a Xen guest OS could be self-paging, it also had to be possible for it to be self-migrating, able to migrate its state to another host, without outside help. After a few months of experimentation, the first prototype implementation of a self-migrating operating system was functioning, and while my research has since strayed in other—still VM related—directions, self-migration and self-checkpointing have remained at the crux of my efforts. Live VM migration in the form first implemented in NomadBIOS, is now a popular technology, because it allows system administrators to reconfigure computing resources without inconveniencing users. In contrast, the interest in self-migration has been more academic, likely because its design goals— simplicity and security—provide only little benefit in the short-term, as large-scale VM deployments are still too rare to attract any significant hacking attacks. The topic of security has increasingly attracted my …