|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iscsi : iscsi parameter default values
Sanjeev,
On the contrary, Immediate Data is simple, it permits important
performance improvements, and the memory arguments are all old arguments.
Perhaps they were valid in the time that FC was being created, but the
memory arguments are not now valid. (see my, and Julian's previous notes
on this.)
Almost every one I know is able to easily code chained Buffer management,
which permits the zero copy capability, and with that the rest of Immediate
Data handling is very easy since you have the associated command with the
data. Using Immediate Data is the simple path, the conservative path.
ImmediateData=yes should be the default. We should not be bound by the
limitations of previous memory technology.
.
.
.
John L. Hufferd
Senior Technical Staff Member (STSM)
IBM/SSG San Jose Ca
Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688
Home Office (408) 997-6136
Internet address: hufferd@us.ibm.com
"Sanjeev Bhagat (TRIPACE/Zoetermeer)" <sbhagat@tripace.com>@ece.cmu.edu on
09/30/2001 12:50:59 PM
Sent by: owner-ips@ece.cmu.edu
To: "'Robert Snively'" <rsnively@Brocade.COM>, Julian
Satran/Haifa/IBM@IBMIL, ips@ece.cmu.edu
cc:
Subject: RE: iscsi : iscsi parameter default values
All,
I guess this is a very important input from robert.
"In FCP, we found that immediate data caused
sufficient complexity, both in parsing command structures and
in performing error recovery, that we chose not to allow it
since there was no performance benefit anyway."
Considering this i have a question is immediate data really needed at all?
and definitely the default value should be no in this case.
Sanjeev
-----Original Message-----
From: Robert Snively [mailto:rsnively@Brocade.COM]
Sent: Sunday, September 30, 2001 7:48 AM
To: 'Julian Satran'; ips@ece.cmu.edu
Subject: RE: iscsi : iscsi parameter default values
Julian,
Are not both immediate data and "unsolicited" data outside the
command PDU both governed by the first-burst size? If so,
the problem remains the same. If you have 50 initiators through
each of 10 ports to 50 logical units with 50 commands queued in
each with immediate data of 4K bytes, that adds up to a lot of
bytes which have to have buffers pre-reserved to allow normal
operation. It is the buffer management that is invasive, not
the indexing of the buffer contents. It is the dual-copy requirement
that is invasive, significantly increasing buffer utilization.
For these problems, immediate data and unsolicited data are
equivalent. Frankly, I see no functional difference between
sending data in the same PDU and sending data in an immediately
following PDU. In FCP, we found that immediate data caused
sufficient complexity, both in parsing command structures and
in performing error recovery, that we chose not to allow it
since there was no performance benefit anyway.
> -----Original Message-----
> From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
> Sent: Friday, September 28, 2001 10:29 PM
> To: ips@ece.cmu.edu
> Subject: RE: iscsi : iscsi parameter default values
>
>
>
> Robert,
>
> Unlike FCP - iSCSI has two forms of unsolicited "immediate"
> and "separate
> unsolicited". They can be used separately or together.
> Immediate data is a single PDU comming together with the
> command while the
> separate unsolicited may come after.
>
> FCP has only the second type.
>
> With separate unsolicited the target has to have the buffers
> for the burst
> and find the "control blocks" by indexing based on ITT.
>
> With Immediate data the target has to deal with a single PDU (not the
> entire burst). Indexing is not done twice (it is done as a
> check for the
> Control block that is being built).
>
> This is why Immediate Data is considered far less invasive
> than separate
> unsolicited (a single buffer, and no indexing) and give the
> performance
> boost it gives we are going to see it probably on every target.
>
> Julo
>
>
>
>
> Robert Snively
>
> <rsnively@broc To: Julian
> Satran/Haifa/IBM@IBMIL, ips@ece.cmu.edu
> ade.com> cc:
>
> Subject: RE:
> iscsi : iscsi parameter default values
> 29-09-01 01:31
>
> Please respond
>
> to Robert
>
> Snively
>
>
>
>
>
>
>
>
> Julian,
>
> For SCSI Write operations, ImmediateData = yes is
> the most demanding mode for the target. If I understand
> it correctly, it essentially over-rides the R2T state,
> allowing the first burst to be included as immediate data.
> In SCSI, except for special cases like this, the target
> is the device in charge of data transfers.
>
> This mode requires the target to have a guaranteed collection of
> received buffer space adequate to receive all possible
> outbound queued operations from all possible initiators
> through all possible target sessions (which may sum to
> 1000s of output operations), regardless of what other
> buffer-intensive operations may be going on in the device.
>
> It then requires the device to provide association with each
> of those commands instead of just the commands it has elected
> to activate. Without immediate data, the command buffer can be
> separated and separately managed from the data buffer,
> limiting buffer requirements.
>
> It then requires the device to operate in a non-zero-copy
> mode (which raises buffer utilization and increases latency)
> while the device analyzes the command to determine what should
> be done with the data. It further requires the subsequent
> data (if there is more than one PDU to be transmitted) to
> be separately managed, and perhaps actually stored in a
> separate operation if the buffer must be cleared to make
> space available for it. It further requires the target to
> analyze the data for completeness before going on to determine
> what to do with it.
>
> Sure, it is easy for the initiator, but so is everything else
> except out-of-order reads.
> It is the target you are stressing with this.
>
> For local sub-LANs, and depending on the actual buffering
> available in the target, the performance may actually be lower with
> ImmediateData allowed, because zero copy behavior of the
> target to non-volatile media is not available, which raises
> buffer utilization.
>
> Have I missed something that would change my mind?
>
> Bob
>
>
> > -----Original Message-----
> > From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
> > Sent: Friday, September 28, 2001 10:38 AM
> > To: ips@ece.cmu.edu
> > Subject: RE: iscsi : iscsi parameter default values
> >
> >
> >
> >
> > Robert,
> >
> > I disagree that Immediate is "the most demanding" mode. That
> > is not my
> > experience and not what I heard from most of the developers. On the
> > contrary immediate data do not require a separate indexing
> > operation on the
> > target (as non-immediate unsolicited do) and that was the
> base for the
> > consensus to have them negotiated separately.
> >
> > For the software initiator it is the most inexpensive way to
> > perform short
> > data transfers (4-8k) typical for major applications like
> > Database access
> > and so it is for lightweight target.
> >
> > Julo
> >
> >
> >
> >
> >
> > Robert Snively
> >
> > <rsnively@broc To: John
> > Hufferd/San Jose/IBM@IBMUS, Julian
> > ade.com>
> > Satran/Haifa/IBM@IBMIL
> > cc:
> > ips@ece.cmu.edu
> > 28-09-01 17:25 Subject: RE:
> > iscsi : iscsi parameter default values
> > Please respond
> >
> > to Robert
> >
> > Snively
> >
> >
> >
> >
> >
> >
> >
> >
> > I vote no as the default value on ImmediateData.
> >
> > A default value of yes on ImmediateData requires implementation
> > of the most complex and demanding mode of operation as the
> > default. SCSI has traditionally made its default behavior the
> > simplest and most encompassing, setting more sophisticated
> > behavior by subsequent agreement. While it may be your earnest
> > desire to encourage the implementation of this function, it
> > is not appropriate to demand that as the default behavior
> > of all devices. In particular, there is no special benefit
> > to providing ImmediateData in low-cost local sub-lans.
> >
> > If you want to encourage it in a profile, fine, but demanding
> > it as the default in the core standard is not appropriate.
> >
> > Note that the behavior of SCSI is traditionally managed
> > entirely by the target. As such, there has never before now
> > been a requirement for the target to, as a default, accept
> > any PDU except a command or task management function
> > that was not explicitly solicited. That is one of the mechanisms
> > that assists SCSI in achieving a low-overhead zero copy
> > capability while operating with a large number of initiators
> > and with deep command queues.
> >
> > Bob Snively e-mail: rsnively@brocade.com
> > Brocade Communications Systems phone: 408 487 8135
> > 1745 Technology Drive
> > San Jose, CA 95110
> >
> >
> >
> > > -----Original Message-----
> > > From: John Hufferd [mailto:hufferd@us.ibm.com]
> > > Sent: Friday, September 28, 2001 1:33 AM
> > > To: Julian Satran
> > > Cc: ips@ece.cmu.edu
> > > Subject: RE: iscsi : iscsi parameter default values
> > >
> > >
> > >
> > > I also agree with this. It should be yes.
> > >
> > > .
> > > .
> > > .
> > > John L. Hufferd
> > > Senior Technical Staff Member (STSM)
> > > IBM/SSG San Jose Ca
> > > Main Office (408) 256-0403, Tie: 276-0403, eFax: (408) 904-4688
> > > Home Office (408) 997-6136
> > > Internet address: hufferd@us.ibm.com
> > >
> > >
> > > Julian Satran/Haifa/IBM@IBMIL@ece.cmu.edu on 09/27/2001
> 09:50:21 AM
> > >
> > > Sent by: owner-ips@ece.cmu.edu
> > >
> > >
> > > To: ips@ece.cmu.edu
> > > cc:
> > > Subject: RE: iscsi : iscsi parameter default values
> > >
> > >
> > >
> > >
> > > The one that I have trouble living with is ImmediateData.
> > > This is important
> > > for low-end desktops and hardly matters for large boxes.
> > > As such I would suggest it stays as yes.
> > >
> > > Julo
> > >
> > >
> > >
> > > "Eddy
> > > Quicksall" To: "'Santosh Rao'"
> > > <santoshr@cup.hp.com>,
> > > <EQuicksall@med <ips@ece.cmu.edu>
> > > iaone.net> cc:
> > > Sent by: Subject: RE:
> > > iscsi : iscsi
> > > parameter default values
> > > owner-ips@ece.c
> > > mu.edu
> > >
> > >
> > > 27-09-01 17:22
> > > Please respond
> > > to "Eddy
> > > Quicksall"
> > >
> > >
> > >
> > >
> > >
> > > I like your defaults below.
> > >
> > > But, section 5 says:
> > >
> > > The initial Login request MUST include the InitiatorName and
> > > SessionType key=value pairs.
> > >
> > > Since SessionType is REQUIRED, naming a default would imply a
> > > possible typo
> > > in the spec.
> > >
> > > Eddy
> > >
> > > -----Original Message-----
> > > From: Santosh Rao [mailto:santoshr@cup.hp.com]
> > > Sent: Wednesday, September 26, 2001 10:29 PM
> > > To: ips@ece.cmu.edu
> > > Subject: iscsi : iscsi parameter default values
> > >
> > >
> > > All,
> > >
> > > With the issue of mode page vs. login keys having [almost]
> > drawn to a
> > > close, I would like to
> > > raise the below issues again, on the subject of default
> > > values for login
> > > keys and other iscsi
> > > parameters :
> > >
> > >
> > > * In keeping with traditional use of scsi mode pages,
> > iscsi should
> > > not specify any default
> > > settings for any mode pages that continue to exist for iscsi.
> > > "Default values" for mode
> > > pages are target specific and should not be bound to
> > the protocol
> > > draft.
> > >
> > > * MORE IMPORTANTLY, I read the default for EMDP as
> being set to 1
> > > :-< This implies that
> > > targets must be always prepared to deal with out of
> > > order data and
> > > initiators must be
> > > prepared to deal with out of order data, unless the initiator
> > > performs a mode select to
> > > disable it. [which defeats all the previous advantages
> > > gained thru
> > > mandating use of login
> > > keys for other negotiations.]. A default, if it were
> to exist,
> > > should be 0. (in-order, by
> > > default).
> > >
> > > * Conservative specification of defaults for login
> keys along the
> > > following lines :
> > > MaxConnections = 1
> > > FMarker = "no"
> > > InitialR2T = "yes"
> > > BidiInitialR2T = "yes"
> > > ImmediateData = "no"
> > > DataPDULength = 16
> > > MaxOutstandingR2T = 1
> > > DataPDUInOrder = "yes"
> > > ErrorRecoveryLevel = 0
> > > SessionType = "normal"
> > >
> > > * Should the iscsi protocol require a "Lun Control Mode
> > Page"? IOW,
> > > is an EnableCRN bit
> > > required at the transport layer ? If the device server
> > capability
> > > is to be negotiated , I
> > > suggest this bit be moved to a SCSI ULP Mode Page such as the
> > > "Control Mode Page", through a
> > > T10 change as a part of the SCSI changes being
> driven by iscsi.
> > >
> > > Comments ?
> > >
> > > Thanks,
> > > Santosh
> > >
> > >
> > > Santosh Rao wrote:
> > >
> > > > There are the separate issues of :
> > > >
> > > > * iscsi's specification of defaults for its mode pages
> > > which is not
> > > in line with mode page
> > > > usage. This impacts the target's ability to enforce
> > > values other
> > > than the protocol
> > > > specified default, if the initiator were to not use mode
> > > sense/select.
> > > >
> > > > * default settings for login keys.
> > > >
> > > > * Is there a need for the "LUN Control Mode Page" and whether
> > > "Enable CRN" should be in a
> > > > transport specific mode page ?
> > > >
> > > > which need to be driven to closure as well.
> > > >
> > > > Regards,
> > > > Santosh
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> >
> >
> >
> >
> >
>
>
>
>
>
Home Last updated: Mon Oct 01 09:17:18 2001 6907 messages in chronological order |