Hi Scott,

On Mon, 5 Nov 2001 16:26:25 -0500
Scott Young <scottyoung at adelphia.net> wrote:

> On Monday 05 November 2001 03:42 pm, you wrote:
> > On Monday 05 November 2001 08:15,  toad wrote:
> > > On Mon, Nov 05, 2001 at 03:40:05AM -0500, Gianni Johansson wrote:
> > >...
> > >
> > > Could you please make sure this uses full blocks (i.e. PO2 sizes) for all
> > > the chunks (as many of them as possible anyway)?
> >
> > It already does this. Blocks are 256k, 512K or 1Mb depending on the file
> > size.   The final block may be zero padded if nescessary.
> 
> Full blocks might not always be the best way to insert data.  Let's say you 
> have file B which is an updated version of A.  B might have its data shifted 
> slightly, so its blocks might not align, creating unnecessary redundency if 
> inserted into freenet even with the same block size.  Upon inserting B, what 
> if FProxy could use the checksums on the segments of A, then calculate the 
> checksums at all offsets in B, and create the splitfile index of B using the 
> blocks on A that match B, and just insert blocks that fill in the changes and 
> the missing data?  This would increashttp://samba.org/~tridge/phd_thesis.pdfe 
> reliability of the overlapping blocks, 
> and reduce the total file size on freenet.  The only problem with this is 
> that it would make some blocks smaller than others, which would make FEC more 
> complicated.  Maybe an FEC algorithm that can work well with variable-size 
> blocks can be used, and maybe it could have some abilities for use with 
> overlapping blocks.  Or for simplicity with FEC, the partial blocks could be 
> padded and/or combined with other partial blocks and somehow be specially 
> indexed in the splitfiles' indexes.

Something like this has been done. Perhaps you will find this link useful:

http://samba.org/~tridge/phd_thesis.pdf

Its Andrew Tridgell's PhD thesis. One of the chapters is about Rsync.
Rsync uses a rolling CRC and a block CRC to allow just the sort of thing
you describe, but in a different context. Still, it may be interesting
reading for you.

Keep well,
Ray



_______________________________________________
Devl mailing list
Devl at freenetproject.org
http://lists.freenetproject.org/mailman/listinfo/devl

Reply via email to