I have this problem described in FAQ:
it is possible for one connection or a small number of connections to take
all the bandwidth from a given bucket and the other connections to be
starved completely, which can be a major problem if there are a number of
large objects being transferred and the parameters are set in a way that a
few large objects will cause all clients to be starved (potentially fixed by
a currently experimental patch).

I have a delay pool set for downloading files ie. exe,mp3 ...
When some people start to use download accelerators and open 30 connections
per file or so
a normal user simply can't download anything.

I have the latest stable version 2.5stable5. Can't find this experimental
patch that should solve this problem. Anyone know where to get it?