SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


    [Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

    RE: iSCSI: SCSI Iniitator port: a simple approach, acceptable?



    Jim,
    
    Multiple connections, especially when spread across multiple
    adapters will give performance that will be much less than
    that of multiple sessions in parallel (of course when there
    are multiple connections/session over multiple adapters - the
    adapters have to be able to take the session id from the
    host software - I obviously don't get the point of the debate).
    
    There is no way a common synchronization point for
    command sequencing and flow control at the host level will not
    cause sub-standard performance. I have argued this point
    many times before, only to have it be brushed aside. Of course,
    only time will tell.
    
    Somesh
    
    > -----Original Message-----
    > From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of
    > Jim Hafner
    > Sent: Thursday, October 11, 2001 2:34 PM
    > To: Santosh Rao; ips@ece.cmu.edu
    > Subject: Re: iSCSI: SCSI Iniitator port: a simple approach, acceptable?
    >
    >
    >
    > Santosh,
    >
    > I agree we haven't really solved the problem exactly (about what has to be
    > configured per HBA, though I hinted at an implementation of portal group
    > tag assignment that is based on network stuff, that might make it easier.
    >
    > It's your last comment that I was expecting to get solicited.
    > But I'll ask
    > why is it so important to have several sessions in parallel?
    > If I can have multiple connections per session, doesn't that suffice for
    > multiplexing?  If the target says only one connection per session, then
    > isn't that a hint that it has limited network resources and it
    > would be bad
    > form to build too many sessions just to bypass this?
    >
    > I'm not strongly advocating this, playing devil's advocate.
    >
    > Jim Hafner
    >
    >
    > Santosh Rao <santoshr@cup.hp.com>@cup.hp.com on 10/11/2001 02:10:56 pm
    >
    > Sent by:  santoshr@cup.hp.com
    >
    >
    > To:   Jim Hafner/Almaden/IBM@IBMUS
    > cc:   ips@ece.cmu.edu
    > Subject:  Re: iSCSI: SCSI Iniitator port: a simple approach, acceptable?
    >
    >
    >
    > Jim,
    >
    > I am not in favor of the below proposal and would rather see the ISID
    > model, for the reason you mention under "Cons".
    >
    > Up until now, the concept of "initiator portal groups" has not held any
    > significant importance in the protocol, unlike target portal groups
    > which is formal information exchanged b/n initiator and target, and
    > which governs the session end points.
    >
    > How is the initiator portal group now going to be communicated to the
    > iscsi HBAs ? In essence, we've just moved the piece of information that
    > needs to be co-ordinated across HBAs, from being an ISID to being a
    > portal group tag.
    >
    > For single connection sessions, it is essential that several sessions be
    > allowed in parallel b/n portal groups.
    >
    > Regards,
    > Santosh
    >
    >
    > Jim Hafner wrote:
    > >
    > > Folks,
    > >
    > > I'm curious to know from this group whether the following is an
    > acceptable
    > > alternative to the complex issues we're battling with the definition and
    > > behavior of SCSI Initiator Port (viz a viz, ISID, key=value,etc.).
    > >
    > > Make the SCSI Initiator Port = initiator portal group, named by iSCSI
    > Name
    > > + portal group tag (which would be presented during login).
    > >
    > > Pros:
    > > a) the model is completely symmetric on the initiator and the
    > target side
    > > b) namespace is managed by any mechanism to identify the portal group
    > tag.
    > > (For example, the tag could be the position in the SCSI port/module load
    > > sequence (/dev/scsi[x]) OR one could derive the portal group tag from
    > > either a MAC or from ipaddress of one of the interfaces in the portal
    > > group).
    > > c) the drivers present exactly one channel (SCSI initiator
    > port) for each
    > > initiator portal group to the upper layers in the OS (how that is
    > > named/identified to the upper layers may be independent of how it is
    > named
    > > on the wire, i.e., independent of the portal group tag).
    > > d) completely static/stable configuration on each boot (though it may
    > > change, as it can today, between boots if components fail or are changed
    > in
    > > any way).
    > >
    > > Cons:
    > > a)  we *prohibit more than one session* between an initiator
    > portal group
    > > and a target portal group (to avoid parallel nexus).  This has the side
    > > consequence of making it harder to be a bad net-citizen by opening too
    > many
    > > connections through the same set of network interfaces (you
    > open what you
    > > want within a session only).
    > >
    > > I've always avoided this approach as I expected the "cons" would not be
    > > acceptable, but I've never actually asked this of the group.
    > >
    > > So, what's the 'acceptability' of this approach?
    > >
    > > Jim Hafner
    > >
    > > Jim Hafner/Almaden/IBM@IBMUS@ece.cmu.edu on 10/11/2001 08:58:41 am
    > >
    > > Sent by:  owner-ips@ece.cmu.edu
    > >
    > > To:   Black_David@emc.com, ips@ece.cmu.edu
    > > cc:
    > > Subject:  Re: iSCSI: ISID progress
    > >
    > > David,
    > >
    > > More comments in line
    > >
    > > Jim Hafner
    > >
    > > Black_David@emc.com on 10/10/2001 03:48:55 pm
    > >
    > > To:   Jim Hafner/Almaden/IBM@IBMUS, ips@ece.cmu.edu
    > > cc:
    > > Subject:  iSCSI: ISID progress
    > >
    > > This message is shorter than my last one, so that's
    > > at least a superficial indication of progress ;-).
    > >
    > > Jim and Julian had four comments/objections to the
    > > notion of using a text key to indicate iSCSI Initiator
    > > Port Name.  Summarizing and responding:
    > >
    > > Jim (a): Optional use of port name is ok as far as SAM-2 is
    > >      concerned, but is odd.
    > >
    > > Indeed it is odd, but given the choice between an odd
    > > mapping of SAM-2 to iSCSI and allowing odd behavior of
    > > iSCSI implementations, I'll take the former.
    > > <JLH>
    > > I'm not sure what "odd behavior of iSCSI implementations" you're
    > referring
    > > to here.
    > > </JLH>
    > >
    > > Jim (b): Would require at least as much coordination above
    > >      the session level of iSCSI as ISIDs.
    > >
    > > That would be incorrect.  128 bits is sufficient to eliminate
    > > coordination.  The reference for this is an expired Internet-
    > > Draft, draft-leach-uuids-guids-01.txt, that can still be found
    > > on the web at:
    > >
    > > http://casbah.org/cbRFC/misc/draft-leach-uuids-guids-01.txt
    > > http://globecom.net/ietf/draft/draft-leach-uuids-guids-01.html
    > >
    > > I'm not seriously proposing that port name generation be done
    > > in this fashion, but rather providing a widely used counter-
    > > example to Jim's statement.  Note that a network interface
    > > MAC is likely to be available to many iSCSI implementations.
    > > <JLH>
    > > I glanced through that draft and certainly don't think it is the right
    > > approach to this problem.  However, if you/we believe that by simply
    > making
    > > the equivalent of the ISID field or portname text value big enough, we
    > can
    > > find a solution to the "coordination problem", why can't we just
    > *require*
    > > that sessions have a SCSI port name (extension to iSCSI Name) that is
    > long
    > > enough to solve the weak or no coordination problem, instead of making
    > this
    > > "TBD"?
    > >
    > > I've argued that having the SCSI Portname is valuable.  I've argued that
    > > there wasn't any net value in putting the name in a key because
    > I already
    > > had the ISID field as part of the "name space for the session endpoint".
    > > But if the claim is simply the ISID isn't big enough, then I
    > don't have a
    > > problem with effectively making them bigger (either in the
    > header or in a
    > > key value).   I've even suggested one way to do that by embedding the
    > > initiator portal group tag in the value.  But another way is to
    > embed the
    > > OUI of the HBA (except for the cases where there isn't an HBA) or embed
    > one
    > > of the MACs of one of the nics (except for the cases where there isn't a
    > > nic, e.g., dialup), etc.  But we'd have to solve all the possible cases
    > > (all of which the NDT had to deal with for iSCSI Names in the first
    > place).
    > >
    > > So, are we at the point where (for this alternative proposal):
    > > (a) we just need a bigger (than 2byte ISID) field for the SCSI portname
    > > extension
    > > (b) we just need an algorithm that lets each component that wants to
    > > generate such a name can do so without collision concerns?
    > > </JLH>
    > >
    > > Jim (c): How to describe model when the text key is optional?
    > >      "Is it that all initiator session endpoints that don't
    > >       provide this text key have *implicit* unique names and
    > >       only when the text key is presented does the name get
    > >       explicit (and then possibly not be unique)? In that case,
    > >       the key would have to be supplied in login next to the
    > >       InitiatorName.
    > >
    > > Yes and yes when it's used, in that order.  Jim's comment (a)
    > > about this being odd applies.
    > >
    > > Julian: ... and have as much chances to blow a session as ISID
    > >
    > > That would also be incorrect.  As previously stated, ISID conflict
    > > is fatal to one of the sessions involved (one cannot change the
    > > ISID and continue), and can occur in any system that opens parallel
    > > sessions.  This text key conflict need not be fatal (one can
    > > change the text key and continue negotiation) and can only occur
    > > in systems that want to use the new port-spanning persistent
    > > reservation functionality, as other systems won't use the text
    > > key.  Also see the draft noted in response to Jim (b) above;
    > > Julian's "have as much chance" language is incorrect.
    > > <JLH>
    > > I think this is somewhat misleading about the cost ("fatal").  If we
    > could
    > > find a reasonable model for "reject: ISID in use" (rather than
    > "blow away
    > > the old session"), then the "fatal to one of the sessions" is a nit.
    > This
    > > would happen on the first exchange of a connection so starting
    > over isn't
    > > any big deal (we may not even have to require connection drop, just
    > restart
    > > of the initial message with new ISID).  The same would happen with the
    > key
    > > as that would have to be somewhere in the pre-full feature phase (but in
    > > this case it could happen anywhere in that phase).
    > > </JLH>
    > >
    > > As previously noted, I can also deal with requiring conservative
    > > reuse of ISIDs as a means to address this situation.
    > >
    > > Thanks,
    > > --David
    > >
    > > ---------------------------------------------------
    > > David L. Black, Senior Technologist
    > > EMC Corporation, 42 South St., Hopkinton, MA  01748
    > > +1 (508) 435-1000 x75140     FAX: +1 (508) 497-8500
    > > black_david@emc.com       Mobile: +1 (978) 394-7754
    > > ---------------------------------------------------
    >
    > --
    > ##################################
    > Santosh Rao
    > Software Design Engineer,
    > HP-UX iSCSI Driver Team,
    > Hewlett Packard, Cupertino.
    > email : santoshr@cup.hp.com
    > Phone : 408-447-3751
    > ##################################
    >
    >
    >
    >
    
    


Home

Last updated: Fri Oct 12 19:17:29 2001
7213 messages in chronological order