Re: [Mldonkey-users] [pango 20030105a] new chunks order

[Pierre Etchemaite]

Le Sun, 05 Jan 2003 20:16:53 +0100, Sven Hartge <address@hidden>
a écrit :

>           "new chunks order": experimental. Try to optimize file
>           completion
>                 time using cost-benefit algorithm. The estimations could
>                 be improved a lot, but the basic algorithm is there.
>                 To allow for experimenting, the new algorithm is only
>                 used when random_order_download is enabled.
> 
> in the ChangeLog, but this explanation is a bit short.
> 
> What are the benefits of this new algorithm?

I'm not sure yet :)

They're many discussions about improving file propagation, how not
downloading in totally random order is bad (including trying to get first
and last chunks in priority, or trying too hard to get rare chunks), how
completing chunks quickly is good, etc.

I modified DonkeyOneFile.compute_size so I could know the amount of missing
bytes in each chunk, and better estimate how many clients I should use for
completing a partial chunk.

The idea (that's certainly wrong, I'm now almost sure), was that since the
whole file is completed when all chunks are, the "work per source" should be
as uniform as possible over all chunks.

This is certainly wrong, because we want to complete some chunks as soon as
possible. And since a chunk is often much smaller than the whole file, 
only the scheduling of the last remaining chunks will determine when the
whole file completes (= completing other chunks a bit sooner or a bit later
doesn't matter).

Instead, I think we should try to use all the available sources for the
smallest number of chunks at a time, within some reasonable limit (no need
to use all clients for the same chunk). Work per source should be as uniform
as possible over those chunks only.

I still don't know how hard we should try to get rare chunks when there's an
opportunity...