Re: Transparent compression in the FS

[Jeff Garzik <jgarzik@pobox.com>]

jw schultz wrote:
>>Consider a simple key-value map, where "$hash $n" is the key and the 
>>value is the block of data.  Only three operations need occur:
>>* if hash exists (highly unlikely!), read and compare w/ raw data
>>* write new block to disk
>>* add single datum to key-value index
> 
> 
> Nicely described, but each block now needs a reference count
> which is incrmented if the raw compare yields a positive and
> decremented when a reference to it receives a write.

Yep.  I'm pondering a sort of a closely-interconnected distributed cache 
for archival storage, so I have to care about refcounts on remote 
servers, too.


> It may help to also have a reverse reference somewhere
> from the block to the hash.

I think I can avoid this...


>>Inode block lists would need to be updated regardless of any collision; 
>>that would be a standard and required part of any VFS interaction. And 
> 
> 
> Under current schemes if i overwrite an already allocated
> block of a file the block list need not be updated unless
> the block is relocated.  But that is a nit.

Yep.  For a networked filesystem, this is less annoying and clients will 
not normally notice any slowdown from this.

	Jeff