Hmm.  This brings to mind a hack which Tim Shepard did at least a
decade ago.  He created an "asynchronous block io" socket protocol
family for use by dump(8).  (ok, you can quit barfing now..).
Basically, the idea to be able to post a zillion read requests at
once, and let the disk deal with them in whatever order it found
convenient.

If I remember correctly, you'd "connect" a socket to a particular
block device, after which "writes" of control blocks to the socket
would queue asynchronous block reads; you'd read back control blocks
indicating which reads had been completed.

Apparently it performed quite well on the systems of the time
(microvaxes running 4.3). Probably would do great on scsi disks with
tagged command queueing and all that..

						- Bill