Alan Dearle

Many systems
supporting orthogonal persistence have been constructed on top of the
abstractions provided by conventional operating systems such as Unix.
These abstractions are not ideal for the construction of persistent systems
since they do not allow the hardware to be used in the most effective
manner. This observation motivated the design of Grasshopper, a new operating
system intended to support persistence as a basic abstraction.

This thesis describes the construction
of a persistent environment in which multiple concurrent computations
can execute and manipulate data stored in a shared address space. Experiments
related to this work were carried out using the Grasshopper operating
system as a framework. The persistent computation environment, known as
the PCE, comprises two major facets of work. The first is a log-structured
persistent store, which is used to hold stable representations of the
state of the PCE. Such states include the persistent data stored within
the shared address space and the meta-data describing the execution state
of the computations. The result is a fully persistent system in which
both data and computations may be recovered in the event of a failure.

As computations exchange information
through the shared address space, causal dependencies are created between
them. To preserve the consistency of the PCE, these dependencies must
be preserved across failure. This is the focus of the second facet of
work, which concerns the design of a new optimistic checkpointing strategy
that guarantees to preserve causal dependencies in the event of a failure.
It allows computations to checkpoint independently and uses an algorithm
described by Johnson and Zwaenepoel to identify recoverable sets of checkpoints
lazily.