Re: Proposal: B_ARRIER (addresses wapbl performance?)

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Tue, Dec 09, 2008 at 05:24:19PM -0500, Thor Lancelot Simon wrote:
> On Tue, Dec 09, 2008 at 09:16:32PM +0100, Manuel Bouyer wrote:
> > On Tue, Dec 09, 2008 at 02:47:50PM -0500, Thor Lancelot Simon wrote:
> > > [...]
> > > Note that the default tag reordering scheme isn't supposed to reorder
> > > even simple-tagged commands,
> > 
> > Are you sure about this ? The way I read SAM-3, all simple task queued
> > between 2 ordered task or head of queue task can complete in any order.
> > Even if the write cache is disabled.
> 
> This depends on the value of QUEUE ALGORITHM MODIFIER in the Control
> mode page.
> 
> See table 247 in SPC3: http://www.t10.org/ftp/t10/drafts/spc3/spc3r23.pdf
> and the descriptive text which immediately follows it (page 285 of the
> PDF file).

thanks, I was not aware of this. Do you know if it exists on older
devices ? If not, we should use the ORDERED tag where appropriate
(or use software barriers as we're doing now).

> 
> If we really think we are using ORDERED tags everywhere we should, then
> we should be setting this bit to 1 in the mode page, which should give
> performance that in no way differs from running with WCE.  Elsewise, the
> SIMPLE tags are reordered for read, but treated as ORDERED for write.
> 
> But this still exposes enough data to the drive that it should be able
> to efficiently clear writes, if they are effectively sorted by the host.
> If they are not, however, with QUEUE ALGORITHM MODIFIER set to 0 the
> overall performance should be considerably worse than with WCE set.
> 
> Perhaps that is what Jason has observed.
> 
> It is unfortunate that the initiator can't create new task sets at
> will.  If the scope of ORDERED tags were only over initiator-defined
> task sets, we'd have exactly what we wanted.

Sure.

-- 
Manuel Bouyer, LIP6, Universite Paris VI.           
Manuel.Bouyer%lip6.fr@localhost
     NetBSD: 26 ans d'experience feront toujours la difference
--