Re: iSCSI/iWARP drafts and flow control

[Caitlin Bestler <cait@asomi.com>]

On Friday, August 15, 2003, at 08:04 PM, Mike Ko wrote:

> As I indicated to Caitlin in an earlier e-mail, since each iSCSI
> control-type PDU from the target to the initiator is associated with a
> StatSN, the ExpStatSN from the initiator can be used to advance the
> "credit" at the target.

Agreed.

>
> The question is for iSCSI control-type PDUs from the initiator to the
> target, what is the mechanism for the target to advance the "credit" at
> the initiator?  For PDUs where the Immediate flag is not set, the CmdSN
> and MaxCmdSN mechanism in iSCSI are already available and the new 
> "credit"
> scheme is not necessary.  The unsolicited SCSI Data-out PDUs are 
> covered
> by FirstBurstLength and again the new "credit" scheme is not needed.  
> The
> "credit" scheme is needed for PDUs where the Immediate flag is set or 
> the
> SNACK Request (note that SNACKs are not needed in iSER).  For these 
> PDUs,
> the target may not be able to advance the "credit" at the initiator
> individually.  One possibility is to wait for the initiator to issue an
> iSCSI control-type PDU with a new CmdSN but without the Immediate flag,
> and if the target responds with an iSCSI control-type PDU with ExpCmdSN
> set to CmdSN + 1, then "credit" at the initiator is restored to its 
> full
> negotiated (or by definition) value.

My proposal is to apply a "leaky bucket" algorithm. "Extra" credits are
granted for each CmdSN credit granted, up to a maximum. Each extra
send decrements a fixed amount from the credits. In a pinch, Nop
commands can be used to advance the CmdSN.

It allows for bursts, and the credit-scoring can be done entirely
when sending a message that needs an "extra" credit.


>
> As Caitlin pointed out, since no one has yet identified a specific need
> for per-session negotiation of these credits, a simple rule enumerated 
> in
> the draft may be sufficient.
>

Well, their negotiated credits in either case. It's just whether
they are negotiated on a per-session basis by target and
initiator, or negotiated on a mailing list on a per-protocol
basis. If the volume of the mailing list negotiation gets
too high, everyone will quickly agree to the per-session
negotiations. ;-)