|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: Addition of CmdSN in Data-out PDUYou can't assume the ITT takes a specific order. Eddy -----Original Message----- From: Sandeep Joshi [mailto:sandeepj@research.bell-labs.com] Sent: Thursday, October 11, 2001 10:30 AM To: BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2) Cc: 'Paul Koning'; ips@ece.cmu.edu Subject: Re: Addition of CmdSN in Data-out PDU If we use the fact that the ITT-ordering of the unsolicited Data-out PDUs is identical to the ITT-ordering of the commands sent *within* a connection, then it should be possible to resolve the Data-out PDUs in a constant-time operation[*1] No hash and no cmdSN is required keep a per-connection chain within the command queue and look at its head. [*1]There are a couple of exceptions, due to the leeway the standard provides the initiator on Data-out PDUs. -Sandeep "BURBRIDGE,MATTHEW (HP-UnitedKingdom,ex2)" wrote: > > Since I started this thread I feel I must at least contribute! > > The reason why I proposed putting CmdSN (actually it should be RefCmdSN) in > the Data-out PDUs was to enable the target to have a faster search to > associate unsolicited data out PDUs with its SCSI Command PDU. Solicited > Data-out PDUs do not require this as they have a Target Task Tag. > > If all Command PDUs were queued then I believe this would work just fine. > However, as Santosh correctly pointed out they are not and without repeating > what he said this mechanism would not work for immediate command PDUs. > > I am sure that particular implementations could make this work but the > underlying argument is that it needs to work and be useful to all > implementations. The only benefit I now see of having CmdSN in the data PDU > is as a check as implementations must (and can only) use the initiator task > tag to associated the Data-out PDU with the command PDU. Therefore, IMO it > is not a good enough reason for having CmdSN in the Data-out PDUs simply for > a consistency check. > > The benefit of having the data sent unsolicited to minimise if not eradicate > round trip times far out weighs the overhead in having to perform a search > on receipt of unsolicited data. If we could have developed a well defined > mechanism to overcome this overhead then all well and good and that is what > I attempted. Still, if someone can do this and the solution is simple and > straight forward then I am sure that it will have my backing but until then > ... > > Cheers > > Matthew Burbridge > Senior Development Engineer > NIS-Bristol > Hewlett Packard > Telnet: 312 7010 > E-mail: matthewb@bri.hp.com > > > -----Original Message----- > From: Paul Koning [mailto:pkoning@jlc.net] > Sent: Wednesday, October 10, 2001 5:07 PM > To: Julian_Satran@il.ibm.com > Cc: ips@ece.cmu.edu > Subject: Re: Addition of CmdSN in Data-out PDU > > Excerpt of message (sent 10 October 2001) by Julian Satran: > > Inconsistency can be legitimate. CmdSN is ephemeral. It can be reused, it > > may have large holes and using it in an implementation is as bad as a > > hashed index. > > Not true. > > CmdSN values are sequential, by definition. Yes, clearly there will > be small holes because commands complete out of order. But "large" > holes are unlikely. > > In any case, the target has control over that. I can use an array > whose size is given by the number of pending commands times a > correction factor to account for the likely density of holes. Then > MaxCmdSN would be updated based on two considerations: the ability to > handle more pending commands, and the need to keep the distance > between oldest (lowest) still active CmdSN and MaxCmdSN bounded by the > size of the lookup array. > > So having CmdSN in the DataOut PDU allows this approach, thereby > replacing a hash lookup on a rapidly changing ID space by a simple > array indexing operation. Without CmdSN, you're forced to use a > mechanism that has a lot more overhead (in the insert/remove or in the > lookup, depending on the mechanism chosen). > > paul
Home Last updated: Thu Oct 11 13:17:25 2001 7197 messages in chronological order |