|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: ISCSI: Unsolicited data in draft v12Please have a look at 11-94 - it is all there (plus a slight complication with immediate). Julo
OK, now we are on the same page. Though the example you state, would support your case, it seems really strange for someone using an HBA to have a MaxRecvPDULength greater then the FirstBurstSize. I say that simply because one would normally think that it is the HBA that has the smaller amount of storage, hence a smaller MaxRecvPDULength. However, I can accept that a target might have a very small unsolicited buffer space, but can permit very large Solicited Buffers, and might have an HBA, which has a large buffer on the card, so your example may be possible even with HBAs. You say that you have seen your example at plugfest, and I think that may be the case because most of the Plugfest implementations have been Software Targets, and in those cases, MaxRecevPDULength has nothing to do with the HBA (because there is none), so I think you point may be valid, and is perhaps reasonable in those cases. So even though we had to drift quite a ways, what you were originally proposing was a change of the wordage as follows: In section 2.2.4 You suggested we change "An iSCSI initiator MUST send as unsolicited data either the negotiated amount or all the data if the total amount is less than the negotiated amount for unsolicited data." to something like "An iSCSI initiator MAY choose to send no unsolicited data with a command, or if any unsolicited data is sent it MUST be either the negotiated amount or all the data if the total amount is less than the negotiated amount for unsolicited data." I, for one, have now come off the fence, and support your position, as long as it does not hold up the Draft. Julian what do you think? . . . 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, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "Rod Harrison" <rod.harrison@windriver.com> on 04/10/2002 04:01:03 AM To: John Hufferd/San Jose/IBM@IBMUS cc: "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com>, "'Julian Satran'" <Julian_Satran@il.ibm.com>, <ips@ece.cmu.edu>, <owner-ips@ece.cmu.edu> Subject: RE: ISCSI: Unsolicited data in draft v12 John, I'm confused by your assertion that "FirstBurstSize is always >= (the effective value of) MaxRecvPDULength". For unsolicited data it is true that FirstBurstSize limits the "effective" PDU size, but for solicited data the PDU size is min (MaxBurstSize, MaxRecvPDULength). It seems to be quite common for targets to advertise MaxBurstSize > MaxRecvPDULength > FirstBurstSize. For example we've seen values like MaxBurstSize=256k, MaxRecvPDULength=64k, and FirstBurstSize=16k at plugfests. Assuming unsolicited data are permitted, a SCSI transfer length of say 32k gives the initiator the option of either sending 16k unsolicited, either immediate or as DATA-IN, and then waiting for an R2T for the remainder of the payload, or just sending the command and shipping all the data in a single 32k PDU in response to a single R2T. The target has complete control over R2T mode of course, but when MaxRecvPDULength is greater than the SCSI transfer length it is reasonable to assume the target will issue a single R2T for the whole payload. - Rod -----Original Message----- From: John Hufferd [mailto:hufferd@us.ibm.com] Sent: Wednesday, April 10, 2002 10:42 AM To: Rod Harrison Cc: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2); 'Julian Satran'; ips@ece.cmu.edu; owner-ips@ece.cmu.edu Subject: RE: ISCSI: Unsolicited data in draft v12 Rod, FirstBurstSize is always >= (the effective value of) MaxRecvPDULength (as seen by the target). So I still do not understand your point. MaxRecvPDULength is generally the buffer size on the Chip or HBA, and FirstBurstSize is the total amount of Main Memory reserved for the Unsolicited Data. I used the words, above, "effective value of" since MaxRecvPDULength is a connection value and FirstBurstSize is a Session Value. So it might be that "this" connection could have a really big HBA buffer, one that was bigger then the Declared FirstBurstSize for the Session, however, the Maximum PDU Length is limited by the smaller of the FirstBurstSize or MaxBurstSize that was declared for the Session. Therefore, I called it the effective value of MaxRecvPDULength. However, all of this commentary does not change what I specified above. So is your point only focused on the times when the MaxRecvPDULength is equal to FirstBurstSize? If so, then I do not see the issue. And if equal size is not your point, I am really missing it. . . . 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, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "Rod Harrison" <rod.harrison@windriver.com> on 04/09/2002 05:44:55 PM To: John Hufferd/San Jose/IBM@IBMUS cc: "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com>, "'Julian Satran'" <Julian_Satran@il.ibm.com>, <ips@ece.cmu.edu>, <owner-ips@ece.cmu.edu> Subject: RE: ISCSI: Unsolicited data in draft v12 John, I agree, but this can only happen when FirstBurstSize is greater than MaxRecvPduSize. In that configuration I would expect the initiator to send unsolicited data. - Rod -----Original Message----- From: John Hufferd [mailto:hufferd@us.ibm.com] Sent: Wednesday, April 10, 2002 12:33 AM To: Rod Harrison Cc: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2); 'Julian Satran'; ips@ece.cmu.edu; owner-ips@ece.cmu.edu Subject: RE: ISCSI: Unsolicited data in draft v12 What I thought I said is that it should be just as efficient to send a normal (non immediate) unsolicited data-out PUD, following the immediate data, as it is to wait for an R2T and then send the data-out PDU, and probably a lot more efficient. So I do not see why anyone would do other then that. . . . 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, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "Rod Harrison" <rod.harrison@windriver.com> on 04/09/2002 03:39:15 PM To: John Hufferd/San Jose/IBM@IBMUS cc: "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com>, "'Julian Satran'" <Julian_Satran@il.ibm.com>, <ips@ece.cmu.edu>, <owner-ips@ece.cmu.edu> Subject: RE: ISCSI: Unsolicited data in draft v12 My reading of the new rule was that it covered both immediate and non-immediate unsolicited data. I think though the logic in my example would still apply, there is probably an efficiency gain in sending all the data in single solicited DATA-OUT versus 8k in an unsolicited DATA-OUT and 4k in a solicited DATA-OUT. Just thinking in terms of DMA operations on the initiator HBA I suspect there is an advantage to a single DMA for 12k over one for 8k followed by one for 4k. I am assuming here that the initiator can not speculatively DMA ahead because of buffer space constraints. John, I'm not quite sure what you meant in your second sentence. Are you saying even with the new rule the initiator can choose to not send immediate data even if it has been negotiated to be available? - Rod -----Original Message----- From: John Hufferd [mailto:hufferd@us.ibm.com] Sent: Tuesday, April 09, 2002 11:20 PM To: Rod Harrison Cc: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2); 'Julian Satran'; ips@ece.cmu.edu; owner-ips@ece.cmu.edu Subject: RE: ISCSI: Unsolicited data in draft v12 I understand what you said, however, I thought the issue was for non immediate unsolicited data. The case you made is normal, I think, for the choice for immediate vrs non immediate (either R2T or other unsolicited data). If the Initiator has agreed to use non immediate unsolicited data then it is not clear, using your example, why one would not send the data via a normal unsolicited Data-Out PDU, when ready. . . . 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, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "Rod Harrison" <rod.harrison@windriver.com> on 04/09/2002 12:28:27 PM To: John Hufferd/San Jose/IBM@IBMUS, "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com> cc: "'Julian Satran'" <Julian_Satran@il.ibm.com>, <ips@ece.cmu.edu>, <owner-ips@ece.cmu.edu> Subject: RE: ISCSI: Unsolicited data in draft v12 As John has guessed my thinking, at least in part, was of dealing with congestion in the HBA. I can imagine there might be times when the initiator receives a command and cannot provide buffer space to accommodate the unsolicited data, in which case it might be able to ship the command and deal with the data when the target sends R2Ts. I also think there might be efficiency gains to be had by allowing the initiator to ship commands without immediate data in some instances. Consider the case where MaxRecvPduSize for the target is 64k and FirstBurstSize is 8k, we've seen this sort of thing quite a bit at plugfests. If the initiator receives a SCSI command with a transfer length of 12k it is probably more efficient to ship the command immediately and then send the whole payload in one DATA-OUT in response to a single R2T from the target, than to ship 8k of immediate data and then 4k in response to an R2T. In fact for any SCSI transfer length greater than FirstBurstSize and less than MaxRecvPduSize this is probably true. I response to Mathew's concern about pre-allocated buffer space being wasted if unsolicited data isn't sent, I believe this might be more of a theoretical concern than a practical one. Any buffer space that is set aside for unsolicited data that can't be used for anything else will be wasted on every command where the SCSI transfer length is greater than FurstBurstSize when the task moves into "R2T mode" anyway. For large transfers that could be a significant proportion of the time. Consider a modest 12 MB transfer with a generous FurstBurstSize of 1MB, that leaves 11 MB of payload to be transferred under the auspices of R2T, during which time the unsolicited buffers are unavailable. Multiply that by the command window size, and then by the number of initiators the target might service and you quickly end up with an impossible situation. - Rod -----Original Message----- From: John Hufferd [mailto:hufferd@us.ibm.com] Sent: Tuesday, April 09, 2002 12:41 AM To: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2) Cc: 'Julian Satran'; BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2); ips@ece.cmu.edu; owner-ips@ece.cmu.edu; Rod Harrison Subject: RE: ISCSI: Unsolicited data in draft v12 Matthew, I would have thought that if there is some special buffer space set aside for the session, whether physical set aside or as a high/low water mark, it would still be available for other tasks in the session, even if some tasks do not use it, so I fail to see the true impact. Perhaps you have seen something or fear something that I do not understand about why a Initiator would negotiate the unsolicited buffer space (FirstBustSize) and then not use it, except for when it had some kind of congestion, or the like. If you state why you think this would happen, perhaps those persons (Rod) that want this "MUST" changed to "MAY", should state why they think it is important to them. I actually do not see the point of either side. . . . 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, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com> on 04/08/2002 03:44:47 PM To: "'Julian Satran'" <Julian_Satran@il.ibm.com>, "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com> cc: ips@ece.cmu.edu, John Hufferd/San Jose/IBM@IBMUS, "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com>, owner-ips@ece.cmu.edu, Rod Harrison <rod.harrison@windriver.com> Subject: RE: ISCSI: Unsolicited data in draft v12 It would not necessarily need separate buffers but it does need to keep some buffers pre-allocated for unsolicited data so when the data arrives unsolicited there is a buffer available in which to place the data. Matthew -----Original Message----- From: Julian Satran [mailto:Julian_Satran@il.ibm.com] Sent: Monday, April 08, 2002 12:00 PM To: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2) Cc: ips@ece.cmu.edu; 'John Hufferd'; BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2); owner-ips@ece.cmu.edu; Rod Harrison Subject: RE: ISCSI: Unsolicited data in draft v12 I am with John here (the third guy that is right) - why would an implementer have separate buffers for solicited and unsolicited data? Julo "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com> 08-04-02 21:43 Please respond to "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" To: John Hufferd/San Jose/IBM@IBMUS, "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com> cc: Julian Satran/Haifa/IBM@IBMIL, Rod Harrison <rod.harrison@windriver.com>, ips@ece.cmu.edu, owner-ips@ece.cmu.edu Subject: RE: ISCSI: Unsolicited data in draft v12 John, It's not so much an implementation problem but one resource management problem in that if unsolicited data has been negotiated then target MUST pre-allocate buffers with which to store the unsolicited when it arrives. The target implementors will decided whether they want to use unsolicted data and take the buffer resource hit in doing so. However, if they do wish to take this hit but the initators decide not to use unsolicited data (even though they have negotiated to use it) then there is potientially a lot of valuable buffer resources tied in up for unsolicited data but which is not being used. Matthew -----Original Message----- From: John Hufferd [mailto:hufferd@us.ibm.com] Sent: Monday, April 08, 2002 11:13 AM To: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2) Cc: 'Julian Satran'; Rod Harrison; ips@ece.cmu.edu; owner-ips@ece.cmu.edu Subject: RE: ISCSI: Unsolicited data in draft v12 Please explain, why an initiator deciding to not send unsolicited data for a specific command causes an implementation problem. That was not clear from your statements. You still need the R2T capability, so what is lost? . . . 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, Cell: (408) 499-9702 Internet address: hufferd@us.ibm.com "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" <matthew_burbridge@hp.com> @ece.cmu.edu on 04/08/2002 10:25:55 AM Sent by: owner-ips@ece.cmu.edu To: "'Julian Satran'" <Julian_Satran@il.ibm.com>, Rod Harrison <rod.harrison@windriver.com> cc: ips@ece.cmu.edu, owner-ips@ece.cmu.edu Subject: RE: ISCSI: Unsolicited data in draft v12 I must express my concern on this issue. From a target point of view once it has negoiated the use of unsolicited data it has to allocate buffer space for that unsolicited data. Now depending on the various parameters this may be a sizeable chunk of valuable resources which it is making available. Now if the decision to use unsolicited data is being moved from a per session to per task basis (which is what this change effectively does) then it puts an awful lot of resource overhead on the target which may never be used. For the reasons above I propose that we do not relax the v12 restriction and keep it as: "An iSCSI initiator MUST send as unsolicited data either the negotiated amount or all the data if the total amount is less than the negotiated amount for unsolicited data." Matthew Burbridge Principal Engineer NSAS-Bristol Hewlett Packard -----Original Message----- From: Julian Satran [mailto:Julian_Satran@il.ibm.com] Sent: Monday, April 08, 2002 9:36 AM To: Rod Harrison Cc: ips@ece.cmu.edu; owner-ips@ece.cmu.edu Subject: Re: ISCSI: Unsolicited data in draft v12 OK - Julo "Rod Harrison" <rod.harrison@windriver.com> Sent by: owner-ips@ece.cmu.edu 08-04-02 14:52 Please respond to "Rod Harrison" To: <ips@ece.cmu.edu> cc: Subject: ISCSI: Unsolicited data in draft v12 I propose we slightly relax the new restriction in draft v12 that the initiator MUST send the maximum permissible amount of unsolicited data. I suggest we change the rule to allow the initiator to either send no unsolicited data, or the maximum permissible. There is no difficulty for the target here since the lack of unsolicited data will be clearly indicated by a command PDU with F bit set and dataSegLen=0. The target will have all the information it needs to immediately issue R2Ts as appropriate. I believe the initiator should be able to make a policy decision on which individual commands should be sent with unsolicited data and which should not. In draft 11.91 section 2.2.4 I suggest we change "An iSCSI initiator MUST send as unsolicited data either the negotiated amount or all the data if the total amount is less than the negotiated amount for unsolicited data." to something like "An iSCSI initiator MAY choose to send no unsolicited data with a command, or if any unsolicited data is sent it MUST be either the negotiated amount or all the data if the total amount is less than the negotiated amount for unsolicited data." - Rod
Home Last updated: Fri Apr 12 04:18:40 2002 9616 messages in chronological order |