Dear Dick, Thank you. Keeping the send buffer mapped upon the sender address space is surely legal. I wonder what will be happening to the sender that has, say, thousands of the outstanding MPI_Isends, all of which mapped their contents to the receiver memory. This will consume kernel memory space and possibly hardware resources. Will this be good for scalability? On a microkernel with so little memory to play with? I'd rather map the send buffer out. As for the send or receive side byte swapping, or, for that matter, any kind of reversible transformation - the receiver will certainly do this once. Note however that proposal #46 talks about collectives as well. If there are several receivers in a Bcast type operation, they will all be doing the transformation instead of the sender doing this once there and back again. The total expense will be higher for the receiver side transformation if more than two receivers are involved. Another couple of scenarios to consider: 3) Imagine send buffers have to pinned in the memory. To avoid doing this too often, these registrations will normally be cached. If more than one send can be used for a buffer or, for that matter, overlapping portions of the same buffer, say by different threads, access to the lookup-and-pin will have to be made atomic. This will further complicate implementation and introduce a potentially costly mutual exclusion primitive into the critical path. 4) I wonder what a const modifier will do for a buffer identifies by MPI_BOTTOM and/or a derived data type, possibly with holes in it. How will this square up with the C language sequence association rules? 5) Note also if both #45 and #46 will be introduced, there will be no way to retract this, even with the help of the MPI_INIT_ASSERTED, should we later decide to introduce assertion like MPI_NO_SEND_BUFFER_READ_ACCESS. The const modifier from #46 will make that syntactically useless. 6) Finally, what will happen in the Fortran interface? With the copy-in/copy-out possibly happening on the MPI subroutine boundary for array sections? If more than one send is allowed, the application can pretty easily exhaust any virtual memory with a couple of long enough vectors. I'm looking forward to your and everybody's opinion on those scenarios. Best regards. Alexander ________________________________ From: mpi-22-bounces_at_[hidden] [mailto:mpi-22-bounces_at_[hidden]] On Behalf Of Richard Treumann Sent: Monday, December 08, 2008 10:56 PM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45) On a single OS under AIX, IBM MPI does map the portion of send side memory that holds the send buffer into the receivers address space. We do this for both contiguous and non-contiguous buffers. The mapping lasts just long enough for the receive side CPU to do a memory copy from send buffer to receive buffer. (see patent 7,392,256) This optimization does not have any effect on the addressability of the send buffer by the sending task. In our case, at least, this optimization does not argue against the proposal. Also, Robert and I had a chat about the byte swap trick and it seems it should be both semantically cleaner and require fewer CPU cycles to do it in the receive buffer. In the receive buffer there is no question that the application must wait for the communication to complete and the swap only needs to be done once (the message flows in with bytes in sender order and the MPI_Recv does one pass of byte swaps if required. In the send buffer trick, the swaps to receiver order must be done and then a second pass is needed to undo it) Dick Dick Treumann - MPI Team IBM Systems & Technology Group Dept X2ZA / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601 Tele (845) 433-7846 Fax (845) 433-8363 mpi-22-bounces_at_[hidden] wrote on 12/08/2008 04:11:45 PM:
[image removed]
Re: [Mpi-22] please review - Send Buffer Access (ticket #45)
Erez Haba
to:
MPI 2.2
12/08/2008 04:13 PM
Sent by:
mpi-22-bounces_at_[hidden]
Please respond to "MPI 2.2"
Dear Alexander,
As far as I recall memory remapping from the main processor to a network device was discussed before (If I recall correctly, in the April meeting). I think that it's close enough to your scenario of remapping to a different process for the purpose of this discussion.
Is your case real? Do you know of systems that do that with MPI? Or is it a hypothetical case?
For the review process, we do need people to review the text; we added this requirement in the last meeting. Regardless, it does not prevent any other person from giving feedback on the proposal. I'm sure that Jeff or Bill can give you're a more formal definition of the review process.
The wiki page state that: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/mpi22/TicketWorkflow To advance to the first reading, a proposal must be reviewed by the lead chapter author and three other reviewers. That review should check the change against the standard text to ensure that the change in context is correct; in addition, the change should be evaluated for completeness. For changes that involve multiple chapters (but are logically related and hence belong in a single ticket), the respective chapter authors must review the changes in their chapters. These reviews must be entered as comments on the ticket. MPI 2.2 Chapter Authors
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Supalov, Alexander Sent: Saturday, December 06, 2008 11:44 AM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45)
Dear Erez,
Thank you. I agree to respectfully disagree on this with you, for the following two reasons:
1) The memory remapping scenario IO brought up a couple of days ago was not discussed before the first voting as far as I can recall. If you have proof to the contrary, I would most kindly ask you to present it. If this cannot be done, I would say that a new issue has been added to the discussion, and we may need to review its substance.
2) Next, I would like to see the definition of the ticket review process that states the reviewers are supposed to only check the proposed text for correspondence with the existing standard. My opinion is that reviewers are there to bring both textual and substantial concerns up when they see the need for this. So far I've been acting on this conviction.
I'm looking forward to the further discussion on the floor.
Best regards.
Alexander
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Erez Haba Sent: Saturday, December 06, 2008 7:45 PM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45) Thanks Alexander,
I respectfully decline your proposal to suspend the review of these tickets. I don't see any specific reference wrt text in your comments; and you don't bring any new issue that has not been discussed before the 1st voting. Thus I don't see any reason to suspend their review.
Thanks, .Erez
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Supalov, Alexander Sent: Friday, December 05, 2008 4:19 PM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45)
Dear Erez,
Thank you. I reviewed the text and found that further polishing of its textual aspects should be suspended until the substance is clarified. I put a comment to this effect into the ticket, as well as into the dependent ticket #46. In my opinion, both tickets are not yet ready to go into the standard and should go into another round of contemplation of their possible repercussions.
In both cases presumed application friendliness is traded for less freedom of implementation. Application developers who disregard the standard now will most likely continue to do this in the future, possibly in some other way. Restricting the freedom of implementation to make their life easier does not seem to be an attractive proposition to me.
If any of the issues identified so far, or comparable issues we cannot fathom at the moment, surface up in one of the future HPC platforms and hinder MPI adoption or transition to MPI-2.2 there, we will have done disservice both to the MPI standard and to the community. I hope this will bear on the minds of those who're going to vote on these two items at the meeting.
Best regards.
Alexander
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Erez Haba Sent: Saturday, December 06, 2008 12:28 AM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45) Dear Alexander,
It is okay and encouraged for people to comment and argue on the proposals. You can add your comments to the ticket arguing your important points. The forum then consider that various points and vote on the proposal.
However for the voting process we need people to review the text and confirm that it does not conflict with the standard and it is reasonable (from language pov) to be included in the standard.
If we are willing to review the text and state that it valid for the standard, that would be great. If you have any comments on the text please send them to me.
Thanks, .Erez
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Supalov, Alexander Sent: Friday, December 05, 2008 2:26 PM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45)
Dear Erez,
Thank you. I'm afraid I would need to have it explained in more detail why review may not include arguments on the substance. If something in the proposal makes one think that the proposed matter may be detrimental to the MPI standard and its implementations, I consider it one's duty to point this out.
Following up on your reply: the segfault situation you described will make an MPI compliant program break. Thus, the implementation will have to keep the send buffer mapped into the sending process address space. This is a limitation on the MPI implementation that should be taken into account during the voting.
Another possibility that has been pointed out earlier was that the proposed change disallows byte swap and other send buffer conversions to be done in place. At least one historical MPI implementation was doing this to a great avail. Who knows what is going to happen in the future?
Best regards.
Alexander
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Erez Haba Sent: Friday, December 05, 2008 11:18 PM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45) I think that the idea is for the reviewers to check the text for any mistakes and compatibility with the existing text, rather than check for the validity of the proposal. The later as I recall is left for the MPI forum assembly.
As for your question, I'm sure that you can answer it yourself. J If the memory is still also mapped to the original process (as with shared memory) that everything is fine. If the memory is removed from the original process, than the app will get an access-violation fault. (if this system works on a page boundary, to take this action it needs to make sure that there are no other allocation on the same page)
Thanks, .Erez
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Supalov, Alexander Sent: Friday, December 05, 2008 2:05 PM To: MPI 2.2 Subject: Re: [Mpi-22] please review - Send Buffer Access (ticket #45)
Hi,
I'd like to review this proposal. Let's consider the following scenario:
- In the MPI_Isend, MPI maps the send buffer into the address space of the receiving process. - In the matching MPI_Recv, the receiving process makes a copy of the mapped send buffer into the receive buffer. - Once the copy is complete, the send buffer is mapped back into the sender address space during the wait/test call.
What will happen one tries to access the send buffer in between?
Best regards.
Alexander
From: mpi-22-bounces_at_[hidden] [mailto:mpi-22- bounces_at_[hidden]] On Behalf Of Erez Haba Sent: Friday, December 05, 2008 10:48 PM To: MPI 2.2 Subject: [Mpi-22] please review - Send Buffer Access (ticket #45) This proposal has passed 1st voting and needs reviewers. We need 3 volunteers to sign-off on this proposal, plus the 3 chapter authors to sign-off on the text.
The Chapter Authors for
Chapter 3: Point-to-Point Communication Chapter 5: Collective Communication Chapter 11: One-Sided Communication
Please add a comment to the ticket saying that you reviewed the proposal, or please send me your comments.
Send Buffer Access: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/45
Thanks, .Erez
--------------------------------------------------------------------- Intel GmbH Dornacher Strasse 1 85622 Feldkirchen/Muenchen Germany Sitz der Gesellschaft: Feldkirchen bei Muenchen Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer Registergericht: Muenchen HRB 47456 Ust.-IdNr. VAT Registration No.: DE129385895 Citibank Frankfurt (BLZ 502 109 00) 600119052
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies. --------------------------------------------------------------------- Intel GmbH Dornacher Strasse 1 85622 Feldkirchen/Muenchen Germany Sitz der Gesellschaft: Feldkirchen bei Muenchen Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer Registergericht: Muenchen HRB 47456 Ust.-IdNr. VAT Registration No.: DE129385895 Citibank Frankfurt (BLZ 502 109 00) 600119052
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies. --------------------------------------------------------------------- Intel GmbH Dornacher Strasse 1 85622 Feldkirchen/Muenchen Germany Sitz der Gesellschaft: Feldkirchen bei Muenchen Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer Registergericht: Muenchen HRB 47456 Ust.-IdNr. VAT Registration No.: DE129385895 Citibank Frankfurt (BLZ 502 109 00) 600119052
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies. --------------------------------------------------------------------- Intel GmbH Dornacher Strasse 1 85622 Feldkirchen/Muenchen Germany Sitz der Gesellschaft: Feldkirchen bei Muenchen Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer Registergericht: Muenchen HRB 47456 Ust.-IdNr. VAT Registration No.: DE129385895 Citibank Frankfurt (BLZ 502 109 00) 600119052
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies. _______________________________________________ mpi-22 mailing list mpi-22_at_[hidden] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22
Intel GmbH Dornacher Strasse 1 85622 Feldkirchen/Muenchen Germany Sitz der Gesellschaft: Feldkirchen bei Muenchen Geschaeftsfuehrer: Douglas Lusk, Peter Gleissner, Hannes Schwaderer Registergericht: Muenchen HRB 47456 Ust.-IdNr. VAT Registration No.: DE129385895 Citibank Frankfurt (BLZ 502 109 00) 600119052 This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies. *