RE: iSCSI: Flow Control

[Pierre Labat <pierre_labat@hp.com>]

Hi,

From the mails on this thread it seems that most
people agree on the following ?

A) Unsolicited data will be allowed by iSCSI.
Well, i think that the protocol must be able to handle
large quantities of unsolicited data to be performant
for the writes on large latency networks.

B) We need of course a flow control on the data
because even if we have a flow control on the commands,
just a few commands with huge unsolicited data could
overwhelm the target buffers.

C) We need a flow control on the commands to avoid
the TASK SET FULL condition. If the target hit this condition
it has to drop command (return task set full) and
drops data associated. Avoiding the task set full condition
avoid too the "dead locks".

D) As noted by several, one TCP connection doesn't "block"
the commands behind the data. The iSCSI adapters (on the
initiator side) must be build in a way to send first the
commands/task management requests. As soon as a command/task mgt
is posted to the adapter, the next iSCSI PDU sent by
the adapter will be the one encapsulating the command/task mgt
request. This has to be true for any command (read/write...)
and task mgt request.


About managing these flow controls:
==================================
The TCP window can handle efficiently the flow control for the data.
Now we need to flow control the commands. The solutions
can be:

1) Use the window described in the draft

Advantage:
---------
 o Flow control the commands

Disavantages:
-------------
 o Some said that as the flow control is per session,
   the target will flow control the whole session
   (several LUs) when the task set of one LU will
   become [almost] full. Hence one LU blocks the other
   ones in the session when they could continue to work.

   This problem can be workaround by the
   target using a right command blocks allocation policy.
   (A command block is a buffer containing the command
    pending (in the task set) in the target).

   Instead of having a policy where the command  blocks
   are allocated per LU, the target could use a policy
   where the command blocks are allocated per session.
   The target maintains a pool of blocks per session not per
   LU. In this case one LU cannot block another one.
   That means that inside a session the balance of the commands
   in the various task sets is driven by the initiator not
   by the target.
   It doesn't seems to be wrong doing that, hence this disavantage
   is NOT a disavantage.

 o If the initiator send a mix of fast and slow commands the slow
   commands can close the command window even if the target have
   slots free for new commands.

2) Replace the window in the iSCSI draft with a command credit.
   The target advertises (with each response) a credit to the initiator.
   The command reference number and status reference numbers
   can be kept for ordering/recovery purpose. Now, the MaxCMdRN
   becomes a credit unrelated to the CmdRn, that's the only change compared
   to the draft. It specifies the maximum number of commands in flight over
   the session.

Advantage
---------
 o flow control commands
 o get rid of the blocking of fast commands by slow commands.

Disavantage
-----------
 o None



3) Use a (iSCSI) credit per LUN (advertised with the command responses)

Advantage
---------
 o flow control the commands (per LUN)

Disavantage
-----------
 o the iSCSI layer on the initiator needs to maintain a state and a flow control
   per LUN



4) T10 specifies a command credit advertising mechanism (credit returned
   with the status + some asynchronous advertisements (see Mike mail/IBA))

Advantage
---------
 o flow control the commands (per LUN)
 o the state of the flow control would be maintained at the SCSI layer,
   it's ok because the SCSI layer already mantains some states/LUN
 o will benefit to FC too

Disavantage
-----------
 o are we able to persuade T10 to specify that quickly?
 o need to change the SCSI layer


Conclusion
==========
It seems to me that with 2) and even with only one TCP connection,
we have a solution very performant in term of throughput, simple,
and with no dead locks.
Do you agree?


Regards,


Pierre