In a persistent computation, the lifetime of data is orthogonal to the method used to access it. This allows a uniform mechanism of manipulating all data. The authors define a persistent operating system with these characteristics: the base abstraction is the persistent object; data are recoverable in consistent form; processes are persistent and resilient; and there is a uniform protection mechanism. They introduce Grasshopper as an example of a persistent operating system.

The authors use the term “container” to mean either a file or the address space of a process, and the term “locus” to mean thread of control. An instantiation of Grasshopper is some ensemble of loci executing in some collection of containers. Only one container at a time is visible to a locus, but many loci may execute in a particular container. Loci can migrate between containers, thereby realizing interprocess communication, and containers can share memory. Protection is afforded by a capability mechanism, and name service is implemented in user-level containers.

Every container is associated with a manager container responsible for data management, fault management, the relationship between physical memory and disk, and the ability to recover the state of the container.

Loci checkpoint their state at will, so Grasshopper can recover data and processes to a self-consistent state. There are situations where it is not desirable to restore state to the last envelope of checkpoints. The state of some loci may be internally inconsistent, the state may lead shortly to a crashed or hung system, or some loci may be cooperating with nonpersistent entities on other machines. I wish the authors had supplied information on how they handle such issues.

The authors demonstrate that they have implemented a functioning Grasshopper by executing a benchmark involving random lookup, full table scan, and record insertion using a 7 Mb database. Comparing their results to an implementation of the same benchmark on OSF Mach, the authors see a 15-times advantage for Grasshopper. Somehow, I doubt it.

The paper is much too short to cover the territory. The authors do not describe any algorithms, not even the critical ones they use to checkpoint loci. They do not explore the implications of checkpointing. It would have been interesting to see a few one-page sketches of user-level containers.