The Locus Distributed Operating System (UCLA, 1983)
Jonathan Ledlie CS 736 March 31, 2000

Instead of harnessing the increaased computing power of netweoked 
machines, LOCUS focusses more on utilizing the additional disks available in a 
networked situation.  The aithors devise a scheme to transparently support 
distributed files while minimally modifying UNIX's file access smeantics.  From 
the users's standpoint, who is supposed to be situated at one of the networked 
machines (or logged into), she can open, create, delete, close, read, write, and 
seek, just like a normal, local file.  The is opposed to the Clusters paper where the 
user is envisioned to be external to the whole system, accessing the machines as a 
group through a single entrypoint.  The authors enforce file integrity, while 
allowing standard smeantics, by distinguishing between who is using the file 
(US), who is storing the file (SS), and who is managing the file (CSS).  Files are put 
into file groups and there is exactly one CSS per filegroup (per partition if the 
system becomes segmented).  Most system calls begin with the user contacting the 
CSS for the location of most recent version of the file.  The storage site is then 
contacted directly for subsequent calls (like read).  File changes are assured 
atomicity through shadow paging, which replaces the (relative) root node of a 
change to commit.  Given the authors' perspective on how distributed machines 
should appear to a user, LOCUS offers a good preliminary solution to the inherent 
difficulties in managing redundancy and shared data.  It offers page replication, 
dynamic migration, shadow paging, and the ability for remote excecution.  They 
leave unresolved byte ordering and security, both of which would be essential ifwe 
were to expand their idea to cover a large heterogeneous network.  Two other ideas I 
had difficulty with were that it appearred that users were still unable to have one 
file span several disks (even if it were in the same file group) and that users were 
able to continue writing when the network became partitioned.  The merging 
required for this seemed so hairy that it might have been better left out entirely.