On 11-05-02 02:40 PM, Jeff Moyer wrote:
..

Given the highly parallel nature of these parts, I wouldn't be surprised
if the ahci queue depth of 31 is one of the main bottlenecks.  Can you
think of a way to extend the ahci driver in this manner to accommodate
devices like this one?

..

I imagine that another performance drain might be libata's lack of
support for host controller queues --> a hardware queuing layer that
exists between the host and the drive's own queuing mechanism.

Most modern ATA/SATA chipsets have something of that nature.
These are designed to eliminate the activity-gaps that happen between
when a drive finishes a command, and when the host interrupt -> BH -> whatever
gets around to submitting a new command.

With commands already batched up in a hardware host queue, the controller
takes care of that part, keeping the drive active more of the time,
increasing IOPS.

Cheers