Re: Command Queue Depth (was asymmetric/Symmetric)

[Stephen Bailey <steph@cs.uchicago.edu>]

> b)  Implement some sort of credit-based mechanism for overflow-avoidance.

ST did this with a technique called `slot accounting'.  Anything that
could be unprocessed in work queue consumes a slot, and there are
various ways to recover the credits.  You could also turn slot
accounting off.

I didn't really groove to the idea, because historically, storage
devices frequently operate in an oversubscribed mode (advertised queue
depth is on a per LUN basis and doesn't account for number of
initiators), and I though the slot accounting mechanism was too rigid
to accomodate this accepted property.

However, as you point out, the SCSI handling of heavy oversubscription
may not be quite right, so perhaps a better solution in iSCSI is
called for.  Certainly the connectivity implied by iSCSI will only
make the problem worse.

Steph