Re: [SSI-users] more information about CFS

[John H. <jo...@Ca...>]

Alceu Rodrigues de Freitas Junior wrote:
> Hello guys,
>
> I'm looking for more information about CFS (Cluster File System) on 
> OpenSSI.
>   
CFS on OpenSSI is a descendant of the CFS system written by Locus for 
the "Transparent Network Computing" product. which was adopted by OSF as 
their network file system.  (It is currently included in HP's 
TruClusters product.)

The biggest difference between CFS and NFS is that CFS implements full 
Posix semantics - file locking, cache coherency and so on.

In terms of implementation CFS consists of two parts:

Simple file operations like remove file, rename and so on are done by 
remote procedure calls from the client machine to the CFS server.

I/O is intimately connected to distributed shared memory - reading a 
file is the process of mapping pages of the file into the memory of the 
node doing the read.  Writing a file is mapping the pages of the file 
into memory and changing them.   At any one moment a page may be 
writable by one node, or readable by many nodes.  Synchronisation 
between nodes is done by exchanging a "token".

For example: say a process on a node wants to read a page of a file.   
The node either has that page in its cache, in which case all is well, 
the data is just copied from the cache into the reading process, or the 
node asks the CFS server for a copy of the page.  The CFS server will 
get the page into it's own cache (if it's not already there) and send a 
copy to the reading node. (*)

Now what happens if a process wants to write a page?  If the page is 
already in the nodes cache, marked writable, all goes as normal.  If not 
the node asks the CFS server for write access to the page.   The CFS 
server will have to tell all people who currently have read access to 
discard their copy, and send a writable copy of the page to the client.

(* Of course I left a step out here - if the page is currently being 
written by some other node it will have to send back its copy, and mark 
it read only).