From: Distributed systems

Despite recent advances in both local and wide-area networking
technologies, network latency is still a major factor in distributed
systems and likely to remain so.  All DSM systems provide some sort of
caching in an attempt to improve the performance beyond that provided
by doing a network access on every reference to a non-local data item.
Each system must decide whether or not to attempt to keep the data
coherent, and, if so, what coherence strategy to use.  The coherence
semantics which may be provided to the programmer include:

- `strict' consistency, where a read always returns the value written
  by the most recent write

- a `loosely' consistent system where the system enforces some form of
  weak consistency guarantees and the application (or compiler or
  user) can indicate synchronisation points where consistency must be
  enforced;

- no automatic consistency mechanism, but provide the user with the
  facilities necessary to implement user level synchronisation and
  consistency.