A friend of mine reminded me of another simple queueing algorithm for
promoting fairness. I forgot the name, but it works with 2 queues: the
active and the inactive queue. Once the active queue starts to be
processed, any new requests that come in are put on the inactive one.
Once the whole queue is processed, active and inactive are swapped.
The new queue would be processed from the other direction, like the
elevator. Maybe this is not even necessary.

While the elevator algorithm bounds the time that a request may stay in
the queue to about (disk size) / (transfer speed), this algorithm bounds
that time to (buffer memory size) / (transfer speed), which is typically
several orders of magnitude lower.

Did anyone try already if this helps sufficiently?
-Olaf.
-- 
___ Olaf 'Rhialto' Seibert - rhialto@       -- Woe betide the one who feels
\X/ polderland.nl  -- remorse without sin - Tom Poes, "Het boze oog", 4444.