From: Distributed systems

Most parallel applications will use some sort of synchronisation
system to order and control accesses to shared data before actually
accessing the data.  The most important thing to note in DSM systems
is that just blindly using standard test and set operations on bytes
in shared pages will produce a high fault rate; faults are usually
expensive, making this approach unacceptable.

Clouds merges locking with the cache consistency protocol, so that the
user may obtain both a lock and the data in one network transaction.
This system has the advantage that no invalidation messages are
required, since the granting of the lock guarantees that there are no
conflicting copies; it has the disadvantage that an explicit
unlock/discard operation is required to release access to the data.
This is acceptable in Clouds, as the DSM system was designed
specifically to support object invocation, so it is easy to discard on
a return.

Munin provides a distributed lock mechanism using `proxy objects' to
reduce network load.  Proxy objects are maintained by a lock server on
each node; when a thread wants to obtain a lock on an object, it
attempts to lock the proxy instead.  The server obtains the global
lock if it is not already held locally.  Global locking is done by
negotiating with all the other lock servers in the system.  Each lock
may be migrated from server to server, and part of the Munin system
allows objects to be migrated along with their locks.

Other systems, such as IVY and Mermaid, use modified versions of classic
multiprocessor synchronisation facilities.