On 9/20/05, Matthew Toseland <toad at amphibian.dyndns.org> wrote:
> On Tue, Sep 20, 2005 at 06:19:35PM -0400, Evan Daniel wrote:
> > On 9/20/05, Matthew Toseland <toad at amphibian.dyndns.org> wrote:
> > > Any ideas for splitfile FEC algorithms apart from the onion code we use
> > > now (vandermonde codes)?
> >
> > Reed-Solomon codes?  They're (iirc) computationally efficient, have
> > flexible settings for redundancy levels, don't care which blocks have
> > errors, and can correct as many missing blocks as there are redundant
> > blocks (or half that number of blocks with errors the decoder doesn't
> > know about).  The relevant research was done in the 1960s, so I highly
> > doubt there are patent issues.
>
> http://www.4i2i.com/reed_solomon_codes.htm
>
> How much redundancy do we need? 0.5 uses 50%, which would probably be
> extremely slow with R/S. Also, it would have to be segmented, just as
> Onion.

You're probably right about this.  R/S is a lot faster at erasures
than errors, though, particularly if we don't even check for errors. 
(192, 128) code would give the same segments as Onion; going to very
large segments would imply 16 bit symbols, which I think would really
kill performance.  Do you have benchmark numbers for Onion?  Also, I
don't know enough about Onion to know whether the memory requirements
are worse or better.  I suspect they're similar.

Evan

Reply via email to