Hi,
Here are a few benchmarks with the rq-dyn patch in. Most differences are
probably due to larger queues. Regressions in Contest io_load, improvement
in OraSim. tiobench shows random IO and streaming write throughput gains
and general improvement in latency.

Right now it appears AS is only really beaten by deadline when using
big TCQ windows. The cause seems to be that deadline will stuff as
many requests into the disk as possible, allowing a massive rescheduling
window, while AS is more selective. Now AS schedules quite well, so I'm
inclined to think that some deadline+TCQ improvements are at the expense
of fairness, although the AS scheduler definitely has room for
improvement.