On Mon, Oct 23, 2000 at 09:40:57PM +0200, Manuel Bouyer wrote:
> On Mon, Oct 23, 2000 at 06:46:28PM +0200, Manuel Bouyer wrote:
> > Hi,
> > I just added tagged queuing support to siop(4), and also reworked a bit the
> > driver so that it should create less interrupt and PCI load.
> > It should also support the QUEUE FULL status (which ncr(4) failed to handle),
> > but I didn't find a drive which would generate such a condition. If someone
> > has the hardware to test please tell me how it goes.
> 
> Don't bother to try it. I've been able to reproduce it at home, by bumping
> the number of tags to 32. It's brocken :) I'm working on it.

I think the "leaky bucket" algorithm Matt proposed is the correct one.  It
is *not* the case that drives always report QUEUE FULL at the same number of
tags, and some drives that occasionally report QF at a small number of tags
really want a large number to perform well.

Of course, the "drives" that *really* want a lot of tags are RAID 
controllers.  For some I/O loads, the performance of the CMD CRD-5440 (just
to pick an old example, but one with which I'm very familiar) scales almost
linearly with the number of tags in use.

-- 
Thor Lancelot Simon	                                      tls@rek.tjls.com
	the effort to perceive simply the cruel radiance of what is