The problem is under investigation. There seems to be a 
problem in the round-robin used for dispatching bandwidth 
between connections. To be clear, when the bandwidth is too 
low, only few connections (always the same ones) get all 
the bandwidth, the other ones are left without transfering 
anything during a long period of time, causing timeouts and 
broken connections.

This problem was added on the TODO list - and should be 
processed "soon" (probably for the 3.32 release)