|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: iSCSI: Out of order commandsJohn, > -----Original Message----- > From: John Hufferd [mailto:hufferd@us.ibm.com] > Sent: Thursday, November 08, 2001 5:19 PM > To: somesh_gupta@silverbacksystems.com > Cc: Robert D. Russell; Julian Satran; ips@ece.cmu.edu > Subject: RE: iSCSI: Out of order commands > > > > I understand your point, however, their are several approaches that I and > others have used to mitigate this issue. We face that problem with Multi > Processing all the time. > > One technique that I have seen use that is similar to what you describe is > to allocate a set of buffers to each processor (in this case an HBA), and > when they are used up, and only then, the go back to the Buffer Mother, to > allocate another Buffer Set. In this way the Processors (HBA) run mostly > independent, but once in awhile get a new supply. There are other > techniques also, but that one clearly works wells and is very simple. I agree with the mechanism above. So here is where the problem is. Let us take a simplification of M adapters and 1 connection per adapter. Let us also assume that the N buffers are given to each adapter. Total number of buffers used - M * N Buffers available to each adapter - N How large of a command window should be advertised in the case of a session wide window. If the window is N, then we are under-utilizing the buffers but we will never drop commands. If we advertise M * N, and the initiator decides to send commands in an unbalanced manner (because 1 connection is working better), then we have a potential problem where the one adapter uses up its allocation - In the meantime the system will provide it some more (but not enough), and there is an imbalance leading to a situation where commands may have to be dropped. (this could also happen if the system did not allocate buffers evenly e.g. in a scenario where the adapters are not using buffers at the same rate). Now we can try to find a balance in the middle, but again there is a loss of efficiency on one axis or the other. > > Therefore, I do not see the Session Window as a problem in the protocol. > > . > . > . > 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 > > > "Somesh Gupta" <somesh_gupta@silverbacksystems.com> on 11/08/2001 04:57:40 > PM > > Please respond to <somesh_gupta@silverbacksystems.com> > > To: John Hufferd/San Jose/IBM@IBMUS > cc: "Robert D. Russell" <rdr@mars.iol.unh.edu>, Julian > Satran/Haifa/IBM@IBMIL, <ips@ece.cmu.edu> > Subject: RE: iSCSI: Out of order commands > > > > I think we may be talking about different things here. > Yes, there are memories that are "on the iSCSI HBA" > (in the case where everything is on a single board, > this would be the memories that are controlled by > only the iSCSI ASIC or whatever) - and these memories > are not shared among multiple HBAs. These may be used > for any book-keeping purposes, contexts, eddy buffering > etc etc by the adapter. > > Then there is memory pool that is owned by the "system". > Adapters will DMA commands/data/responses directly into > and out of this memory pool which is managed by the "system". > However to prevent adapters from running over each other, > and over the "system" processor, producer/consumer > mechanisms have to be implemented. > > This is necessary to maintain order on the system. So > on a long-term basis, the memory in the system is shared > across the HBAs - absolutely yes. But at any given > instance in time, it is partitioned among various > usages. > > So all in all I agree with you. The only point of difference > is that it is not possible that a given buffer (part of > a buffer pool) is accessible to multiple adapters at > the same instance in time. It is a necessary condition to > prevent corruption and chaos. > > > -----Original Message----- > > From: John Hufferd [mailto:hufferd@us.ibm.com] > > Sent: Thursday, November 08, 2001 4:40 PM > > To: somesh_gupta@silverbacksystems.com > > Cc: Robert D. Russell; Julian Satran; ips@ece.cmu.edu > > Subject: RE: iSCSI: Out of order commands > > > > > > > > I guess I disagree with this design approach. The buffers on the iSCSI > > HBAs are for Eddy buffers. Once you have the iSCSI Header > should be able > > to place the data directly into a Shared buffer. If you have a session > > that spans HBAs, that should work just fine. But the buffers match the > > Session Window, and the Buffers need to be shared across the > HBAs, not in > > the HBAs. Again, the HBAs should just have the Eddy Buffers. > > > > . > > . > > . > > 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 > > > > > > "Somesh Gupta" <somesh_gupta@silverbacksystems.com> on 11/08/2001 > 04:31:29 > > PM > > > > Please respond to <somesh_gupta@silverbacksystems.com> > > > > To: John Hufferd/San Jose/IBM@IBMUS > > cc: "Robert D. Russell" <rdr@mars.iol.unh.edu>, Julian > > Satran/Haifa/IBM@IBMIL, <ips@ece.cmu.edu> > > Subject: RE: iSCSI: Out of order commands > > > > > > > > John, > > > > You are correct if all the connections are on > > a single adapter. In this case the buffer pool > > can be shared among all the connections on the > > adapter. > > > > However this is not true if the connections are > > on different adapters (the point being discussed). > > Think of there being a big buffer pool in the board > > into which multiple adapters/chips are "plugged" in. > > However, at any given point in time, these buffers > > are allocated (handed over through command queues or > > whatever) to each of the individual adapters. And this > > is where the problem arises. > > > > The assumption that adapters can share a typical host > > command buffer pool on their own without host intervention > > is incorrect. > > > > Somesh > > > > > -----Original Message----- > > > From: John Hufferd [mailto:hufferd@us.ibm.com] > > > Sent: Thursday, November 08, 2001 4:20 PM > > > To: somesh_gupta@silverbacksystems.com > > > Cc: Robert D. Russell; Julian Satran; ips@ece.cmu.edu > > > Subject: RE: iSCSI: Out of order commands > > > > > > > > > > > > Most folks I know have a buffer that applies to all the > > > connections, not to > > > just one buffer per connection. I fail to see the problem for most > > > implimentations. The window should apply to the session. > > > > > > . > > > . > > > . > > > 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 > > > > > > > > > "Somesh Gupta" <somesh_gupta@silverbacksystems.com>@ece.cmu.edu on > > > 11/08/2001 02:22:30 PM > > > > > > Please respond to <somesh_gupta@silverbacksystems.com> > > > > > > Sent by: owner-ips@ece.cmu.edu > > > > > > > > > To: "Robert D. Russell" <rdr@mars.iol.unh.edu> > > > cc: Julian Satran/Haifa/IBM@IBMIL, <ips@ece.cmu.edu> > > > Subject: RE: iSCSI: Out of order commands > > > > > > > > > > > > Comments below > > > > > > > -----Original Message----- > > > > From: Robert D. Russell [mailto:rdr@mars.iol.unh.edu] > > > > Sent: Thursday, November 08, 2001 11:46 AM > > > > To: Somesh Gupta > > > > Cc: Julian Satran; ips@ece.cmu.edu > > > > Subject: RE: iSCSI: Out of order commands > > > > > > > > > > > > > > > > Somesh: > > > > > > > > > > It seems to me that if a target can only handle x new commands, > > > > > > then its queueing capacity is x, and in the SCSI Response PDU it > > > > > > should set MaxCmdSN = (ExpCmdSn + x - 1), in accordance with the > > > > > > formula in section 2.2.2.1. This in turn controls the number of > > > > > > commands the initiator can send, and thus prevents the incoming > > > > > > commands from overfilling the target's queue. > > > > > > > > > > This is actually a weakness of the multiple connections over > > > > > multiple adapters model (it is not related to ordering). > > > > > The target advertises a command window on a session wide basis. > > > > > At the same time, it has to provide the resource to the adapter > > > > > to be able to DMA those commands to. Since there is no > restriction > > > > > on which connection the commands may be received, either the > > > > > target has to allocate the max resources needed to each adapter > > > > > (thus committing n times the resources actually required), or > > > > > has to "lie" (it would not be a complete lie) which could > > > > > result in running out of resources. One way to fix that would be > > > > > to have the credit on a per connection basis. > > > > > > > > If I understand you correctly, you are saying that deadlock can > > > > occur even if we enforce ordered delivery by initiators and even > > > > if the advertisements are correct and even if both parties obey > > > > the advertisements. > > > > > > I would let Julian enlighten us on that, or whether it just > > > leads to wholesale command drops and retries, or TASK SET FULL > > > is returned (is TASK SET FULL a valid response in this case > > > for a LINKED task when it is not the first command?) > > > > > > > > > > > In other words, doesn't this mean that the standard can lead to a > > > > trap in implementing targets and that, in fact, the advertisements > > > > really should be on a per connection basis? > > > > However, what would be the consequences for initiators if > > > > the advertisements were on a per-connection basis? > > > > > > There is really no issue with adveritisement on a per > > > connection basis. It leads to a protocol change. However, it > > > also allows a much more performant flow of commands. > > > > > > > > > > > > > > > Thanks, > > > > > > > > Bob Russell > > > > InterOperability Lab > > > > University of New Hampshire > > > > rdr@iol.unh.edu > > > > 603-862-3774 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
Home Last updated: Thu Nov 08 21:17:38 2001 7677 messages in chronological order |