Semi-active replication is a variant of active replication in fault-tolerant distributed system design, where during concurrent execution of multiple instances of the same software component, one single 'leader' instance determines the exact data and control flow to be taken by any other. The resulting sequence of determinants is then simulated on any other instance in order to maintain program state consistency.
This talk discusses Xen/VLS, an extension to the Xen Hypervisor presently developed at LRR, targeting semi-active replication for paravirtual machines on commodity x86 platforms. Past applications at the granularity of operating system processes have often been hampered by deep interdependencies between application software and its runtime environment. We describe why system VMs, most notably the paravirtual design approach taken by Xen, have better properties.
Target processors typically lack one fundamental building block assumed by the execution model underlying both leader and follower modes, namely in support of simulating asynchronous events. The system described comprises a substitute based on x86 hardware performance monitoring facilities, with sufficiently similar properties.
Finally, there are a number potential applications beyond high-availability system design. Well-known ones are system debugging or program analysis in network intrusion detection. However, within the core field of system virtual machine design, we believe that there are opportunities for efficient live migration of virtual machines as well.