Re: iscsi : Fragmentation & Reassembly issues in iSCSI.

[julian_satran@il.ibm.com]

answers in text - Julo

Santosh Rao <santoshr@cup.hp.com> on 19/01/2001 02:47:05

Please respond to Santosh Rao <santoshr@cup.hp.com>

To:   IPS Reflector <ips@ece.cmu.edu>
cc:
Subject:  iscsi : Fragmentation & Reassembly issues in iSCSI.




Julian,

Issue :
=====
DataPDULength, PingMaxReplyLength, Fragmentation/Reassembly issues.

Problem :
=======
1) What is the relationship b/n DataPDULength and PingMaxReplyLength ?

If a simple iSCSI test program decided to use the NOP-OUT as a form of
basic testing and began issuing increasing sizes of NOP-OUT, at some
point, it will cross the DataPDULength  PDU size (provided it was able
to negotiate PingMaxReplyLength > DataPDULength, which such a test
program would attempt, in the interests of large I/O testing).

<js> I will introduce in 04 a statement saying that that PingMaxRelayLength
will be limited by the lowest of the two </js>

The DataPDULength introduces fragmentation and reassembly issues at the
iSCSI layer for the NOP-OUT, NOP-IN and Login commands.

(For the SCSI Data PDUs, reassembly is possible based on the Buffer
Offset contained in the data PDU header.)

Question :
========
What drives the fundamental requirement for the DataPDULength ? I see no
benefit for this other than trying to ensure a small data payload in
order to have an effective CRC. Are there any other benefits ?

<js> data multiplexing on a connexion is the other reason. Unsolicited data
are the only case you are not allowed to multiplex in order to avoid
deadlocks. A good resync mechanism, whenever will become available can also
operate effectively only on limited sizes unless you use a full-fledged
RDMA scheme. Everywhere those reasons are no relevant you can set the
maximum sizes to you memory available. And the maximum ULP PDU is not the
TCP MTU.</js>

Is there a need to impose an upper bound (MTU) on the size of PDUs other
than the SCSI Command PDU (containing immediate data) and SCSI Data PDU
?

Solution :
=======
Specify explicitly in Appendix C that DataPDULength is only applicable
for SCSI Command PDU and SCSI Data PDU. The current definition is open
to being mis-interpreted as a form of iSCSI MTU, something that can
result in the iSCSI layer having to deal with fragmentation and
re-assembly issues for non-SCSI PDUs.

<js> What would be those? Text commands? Login? Nops? Having a single limit
seemms simpler </js>
Regards,
Santosh

 - santoshr.vcf