A somewhat unusual mechanism for achieving parallelism in cloud-based systems is explored in this paper. Traditionally, a cloud server would be instantiated on a multiprocessor, and the developer would run multiple processes or threads to take advantage of the underlying hardware. SnowFlock takes this one step further, allowing a developer to easily and efficiently run copies of the whole virtual machine (VM) on other hardware instances.

The challenge is to make this efficient, and SnowFlock does so by using “three key observations. First, it is possible to drastically reduce the time it takes to clone a VM by copying only the critical state and [by] fetching the VM’s memory image efficiently on demand.” Second, it is possible to significantly reduce the VM memory that must be transferred if the operating system knows about new allocations. Third, the similarity of the memory access patterns among “cloned VMs makes it beneficial to distribute [a] VM state using multicast. This allows for the instantiation of a large number of VMs at a (low) cost similar to that of forking a single copy.”

The paper is interesting as an alternative approach to cloud design. It is detailed enough to understand and grasp the concepts without getting lost in the specifics of implementation. It is definitely an important paper for cloud designers.