|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iSCSI: SNACK R2T/Data
Michael,
I had some difficulty understanding your arguments. If you are worried
about references - data can be acked
by just refering to the ITT and sequential number (no TTT as data out don't
need to be acked).
Julo
Michael Schoberg
<michael_schober To: "IPS Reflector (E-mail)"
g@cnt.com> <ips@ece.cmu.edu>
Sent by: cc:
owner-ips@ece.cm Subject: RE: iSCSI: SNACK R2T/Data
u.edu
22-09-01 01:47
Please respond
to Michael
Schoberg
Using [SCSI Command Retry] would be the same as no ACK method and implies
we
could eliminate StatSN as well. Most SCSI devices will work without iSCSI
recovery, but I believe there was consensus that some recovery method be
built-in. Tapes and such won't like unnecessary retries.
Julian's reference for ACK'ing extremely long transfers makes a case for a
Data/R2T ACK (although I think such transactions may be arguably rare). I
still believe that decoupling [Request][Data][R2T] & [Response] is helpful
for building simpler overall designs. By having to know that an ACK refers
to a response which refers to a request which refers to R2T and/or DataSN
is
cumbersome and unnecessary. Involving a second status queue specifically
for R2T/Read Data allows for complete decoupling of the
[Request][Data][R2T]
& [Response]. If we're going to try to recover R2T and ReadData, I'd
prefer
to have it decoupled from other PDUs flowing through the system.
However even if a separate queue is added, Julian's requirement that a
specific ACK request be used is still legitimate for lengthy T->I
transfers.
For that a NOP/ACK type request (simply updating the DataSN/R2TSN queue
pointers) works. ACKing by referencing the original request, in my
opinion,
is still inconvenient.
As it stands now, SNACK isn't broken. Maybe that's the best compromise
we'll see on this issue.
: -----Original Message-----
: From: Sandeep Joshi [mailto:sandeepj@research.bell-labs.com]
: Sent: Friday, September 21, 2001 12:31 PM
: To: ips@ece.cmu.edu
: Subject: Re: iSCSI: SNACK R2T/Data
:
:
:
:
: The ack effect could also be achieved using SCSI Command Retry (use
: the SCSI Command's expDataSN field) The target stops sending
: Read PDUs if it runs out of buffers, and the initiator then
: times out and retries the command. This is what a target
: would have to do in the absence of some data ack.
:
: I looked at the archived emails on this subject (when dataSN
: was called DataRN) The Orlando minutes also have a brief reference.
: I suspect the atmosphere then was that the recovery scheme was
: still partly defined and this DataSN ack scheme was adding to
: the general sense of complexity (..so lets dispose it)
:
: I agree with Julian & Matthew that some form of data acks would
: add value for long transfers. The acks will be cumulative and
: not per ReadData PDU (based on prenegotiated intervals or on
: a "send_ack" bit in the ReadData PDU set by the target)
:
: Btw, I see that the expDataSN field has disappeared from the
: SCSI Command PDU. The Task Management Command also does not
: have anything related to expDataSN. I may have missed an
: email on this ?
:
: -Sandeep
:
: Julian Satran wrote:
: >
: > It does not have to involve a net type but it must be a
: specific request
: > (like SNACK) as there might be no output
: > traffic at the time. The overhead involved is minor (as it
: can be done only
: > once per sequence and be cumulative).
: >
: > Julo
: >
: >
: > Michael Schoberg
: > <michael_schober To:
: "'ips@ece.cmu.edu'" <ips@ece.cmu.edu>
: > g@cnt.com> cc:
: > Sent by: Subject: RE:
: iSCSI: SNACK R2T/Data
: > owner-ips@ece.cm
: > u.edu
: >
: >
: > 21-09-01 18:43
: > Please respond
: > to Michael
: > Schoberg
: >
: >
: >
: > I can see the point of not involving a new ACK request for
: [DATA] & [R2T].
: > However, it would be nice to have something like [CmdSN,
: ExpStatSN][StatSN,
: > ExpCmdSN] for [DATA] & [R2T]. As it stands now, only when the [SCSI
: > Response] is acknowledged for the request that originated
: the [R2T] &
: > [Data]
: > messages can you assume [R2T] & [Data] was successfully
: processed. This
: > means you have hold onto a lot of associations in that
: updating ExpStatSN
: > will ACK more than just status messages.
: >
: > The nice thing about the [CmdSN, ExpStatSN][StatSN,
: ExpCmdSN] mechanism was
: > that it allowed for separate processing queues; status
: messages can be
: > unaware of the command that originated them. This
: disassociation allowed
: > for a simpler implementation (aka hardware assist).
: >
: > One possibility would be to split the [CmdSN,
: ExpStatSN][StatSN, ExpCmdSN]
: > into 16 bit fields rather than 32. Unless there's a
: genuine feeling that
: > more than 65K requests could be outstanding within session,
: I don't see a
: > problem. The extra 32 bits freed up could be used for a
: Data/R2T_SN ACK
: > and
: > would allow a separate queue independent of [CmdSN,
: ExpStatSN][StatSN,
: > ExpCmdSN]. Since each queue doesn't have to maintain state
: knowledge of
: > the
: > other, it allows for simpler design.
: >
: > Just and idea.
: >
: > : -----Original Message-----
: > : From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
: > : Sent: Friday, September 21, 2001 9:41 AM
: > : To: ips@ece.cmu.edu
: > : Subject: Re: iSCSI: SNACK R2T/Data
: > :
: > :
: > :
: > : Santosh,
: > :
: > : None of your arguments makes much sense.
: > : The ACK (a third form of SNACK) can be done at explicit
: > : boundaries and only
: > : by initiators
: > : supporting the within command class. For most of the
: > : initiator it will
: > : require 0 additional resources as they either
: > : don't do they recovery, or the input data is short (the ack
: > : is not needed
: > : at the end as the status ack acks the data too).
: > : For the ones that have long exchanges (tape and 3rd
: party) it is a big
: > : help.
: > :
: > : But you did not fail (again) to make your point.
: > :
: > : Julo
: > :
: > :
: > :
: > :
: > : Santosh Rao
: > :
: > : <santoshr@cup. To:
: > : ips@ece.cmu.edu
: > : hp.com> cc:
: > :
: > : Sent by: Subject: Re:
: > : iSCSI: SNACK R2T/Data
: > : owner-ips@ece.
: > :
: > : cmu.edu
: > :
: > :
: > :
: > :
: > :
: > : 20-09-01 18:54
: > :
: > : Please respond
: > :
: > : to Santosh Rao
: > :
: > :
: > :
: > :
: > :
: > :
: > :
: > :
: > :
: > : Matthew,
: > :
: > : We have gone through this thread of discussion regarding
: > : DataSNa long time
: > : back on
: > : ips and the consensus has been that I/O transfer sizes
: are not large
: > : enough to
: > : justify OUT_OF_BAND acknowledgement of data. [achieved by
: having the
: > : initiator
: > : generate NOP-OUTs in response to received data pdu's to send
: > : a DataSN ack.]
: > :
: > : The primary dis-advantage with that scheme was that the data
: > : ack's were
: > : being
: > : generated out-of-band, and therefore, implied extra cost
: in terms of
: > : initiator
: > : resources for each I/O, as well as increased wire traffic per I/O,
: > : performance
: > : penalty, etc.
: > :
: > : I am opposed to the draft requiring initiators to send
: > : out-of-band ack's to
: > : data
: > : pdu's through the use of initiator generated NOP-OUT
: pdus. This is a
: > : performance
: > : penalty, a resource overhead, and not really justified in
: > : most I/O traffic
: > : due to
: > : the avg. data xfer sizes.
: > :
: > : Regards,
: > : Santosh
: > :
: > :
: > : Julian Satran wrote:
: > :
: > : > Matthew,
: > : >
: > : > I am also in favor of such a mechanism. It is easy to
: > : implement (another
: > : > form of SNACK) and we have already built-in a mechanism
: by which an
: > : inbound
: > : > stream can be marked for acking (the inbound sequence
: > : separator F bit).
: > : > It can also be specified as mandated only if the
: > : within-command recovery
: > : > class is present.
: > : >
: > : > However I am reluctant to open again this issue except if
: > : there are more
: > : > supporters. Data ACKs where hastily dropped at the interim
: > : meeting in
: > : > Orlando. I recall Somesh Gupta, Pierre Labat and Santosh
: > : Rao as being
: > : very
: > : > vocal against them (arguing complexity) and carrying the
: > : room with them.
: > : >
: > : > Julo
: > : >
: > : >
: > : > "BURBRIDGE,MATTH
: > : > EW To: Julian
: > : Satran/Haifa/IBM@IBMIL,
: > : > (HP-UnitedKingdo ips@ece.cmu.edu
: > : > m,ex2)" cc:
: > : > <matthew_burbrid Subject: RE:
: > : iSCSI: SNACK
: > : R2T/Data
: > : > ge@hp.com>
: > : >
: > : > 19-09-01 17:25
: > : > Please respond
: > : > to
: > : > "BURBRIDGE,MATTH
: > : > EW
: > : > (HP-UnitedKingdo
: > : > m,ex2)"
: > : >
: > : >
: > : >
: > : > I am very much in favour of having a positive ACK mechanism
: > : to control
: > : > buffer resources at the target. If there is a very large
: > : transfer (e.g.
: > : 1
: > : > Mb) then the sender can release buffer space once it
: knows that the
: > : > receiver
: > : > has received the data. It is worth pointing out that this
: > : mechanism is
: > : for
: > : > buffer control and is not for flow control which, as we
: all know, is
: > : > handled
: > : > by TCP.
: > : >
: > : > Cheers
: > : >
: > : > Matthew Burbridge
: > : > Senior Development Engineer
: > : > NIS-Bristol
: > : > Hewlett Packard
: > : > Telnet: 312 7010
: > : > E-mail: matthewb@bri.hp.com
: > : >
: > : > -----Original Message-----
: > : > From: Julian Satran [mailto:Julian_Satran@il.ibm.com]
: > : > Sent: Wednesday, September 19, 2001 6:28 AM
: > : > To: ips@ece.cmu.edu
: > : > Subject: Re: iSCSI: SNACK R2T/Data
: > : >
: > : > There is no ACK mechanism. The group wisdom was that there
: > : is no need for
: > : > one.
: > : > Incoming data and R2Ts are not acked (a mechanism that did
: > : that existed
: > : and
: > : > was based on NOP-Out).
: > : >
: > : > Julo
: > : >
: > : > Michael Schoberg <michael_schoberg@cnt.com> on
: 18-09-2001 19:09:51
: > : >
: > : > Please respond to Michael Schoberg <michael_schoberg@cnt.com>
: > : >
: > : > To: "'ips@ece.cmu.edu'" <ips@ece.cmu.edu>, Julian
: > : Satran/Haifa/IBM@IBMIL
: > : > cc:
: > : > Subject: iSCSI: SNACK R2T/Data
: > : >
: > : > Old subject, but I couldn't find any discussion on this:
: > : >
: > : > When does the target know it no longer needs to hold R2T &
: > : Data PDUs?
: > : > StatSN responses are acknowledged through the ExpStatSN
: > : field received in
: > : > future I->T requests. What's the acknowledgement method
: > : for R2T & Data
: > : > PDUs? Is it tied to the original request and acknowledged
: > : through the
: > : > ExpStatSN acknowledgment of the request's response?
: > : >
: > : > Thanks.
: > :
: > : - santoshr.vcf
: > :
: > :
: > :
:
Home Last updated: Sat Sep 22 06:17:24 2001 6676 messages in chronological order |