- mpiwg-coll - mailman3.cels.anl.gov

Re: [mpi3-coll] Next teleconference
by Jesper Larsson Traeff 13 Nov '08

13 Nov '08

dear all, both days, should work for me, strong preference for Wednesday, 11/28 though! Jesper On Wed, Nov 12, 2008 at 09:35:20PM -0500, Torsten Hoefler wrote: > On Wed, Nov 12, 2008 at 08:43:08PM -0500, Torsten Hoefler wrote: > > Can we please do another poll for the week from 11/24 to 11/28? It looks > > like noon EST works pretty well, thus, I would propose: > > > > Wednesday, 11/19 > > Friday, 11/21 > sorry - I actually meant: > > Wednesday, 11/26 > Friday, 11/28 > > All the Best, > Torsten > > -- > bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- > Torsten Hoefler | Postdoctoral Researcher > Open Systems Lab | Indiana University > 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA > Lindley Hall Room 135 | +01 (812) 855-3608 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll

1 0

Re: [mpi3-coll] Next teleconference
by Richard Graham 12 Nov '08

12 Nov '08

11/26 might work 11/28 won¹t work Rich On 11/12/08 9:35 PM, "Torsten Hoefler" <htor(a)cs.indiana.edu> wrote: > On Wed, Nov 12, 2008 at 08:43:08PM -0500, Torsten Hoefler wrote: >> > Can we please do another poll for the week from 11/24 to 11/28? It looks >> > like noon EST works pretty well, thus, I would propose: >> > >> > Wednesday, 11/19 >> > Friday, 11/21 > sorry - I actually meant: > > Wednesday, 11/26 > Friday, 11/28 > > All the Best, > Torsten > > -- > bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- > Torsten Hoefler | Postdoctoral Researcher > Open Systems Lab | Indiana University > 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA > Lindley Hall Room 135 | +01 (812) 855-3608 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll >

1 0

Re: [mpi3-coll] Next teleconference
by Torsten Hoefler 12 Nov '08

12 Nov '08

On Wed, Nov 12, 2008 at 08:43:08PM -0500, Torsten Hoefler wrote: > Can we please do another poll for the week from 11/24 to 11/28? It looks > like noon EST works pretty well, thus, I would propose: > > Wednesday, 11/19 > Friday, 11/21 sorry - I actually meant: Wednesday, 11/26 Friday, 11/28 All the Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

[mpi3-coll] Next teleconference
by Torsten Hoefler 12 Nov '08

12 Nov '08

Hello, I leared from Rich that the Thursday after SC is Thanksgiving. I guess we won't have too many participants if we do it at noon. Can we please do another poll for the week from 11/24 to 11/28? It looks like noon EST works pretty well, thus, I would propose: Wednesday, 11/19 Friday, 11/21 What do you think? New proposals (time,date) are welcome! Thanks & All the Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

[mpi3-coll] Nonblocking collectives standard draft
by Torsten Hoefler 12 Nov '08

12 Nov '08

Hello, as promised in the last forum, I finished the draft of our standard proposal. I don't want to post it to the mailinglist, so I uploaded it to the wiki pages [1] and put a copy at [2] for convenience. The document includes the whole collective chapter and the additions (some references to other chapters are thus broken). All changes are clearly marked: - Striked text has been removed and underlined text has been added. - Section 5.12 is new (not underlined) - Examples 5.28-5.33 are new (not underlined) Please read the additions and send comments to me! I hope to be able to discuss the proposal during the next teleconference (which will be announced in a separate mail). Thanks & All the Best, Torsten [1]: https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/NBColl [2]: http://www.unixer.de/sec/nbc-proposal-11-12-2008.pdf -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

Re: [mpi3-coll] Non-blocking Collectives Proposal Draft
by Torsten Hoefler 09 Nov '08

09 Nov '08

Hi Rich, thanks for your comments! I will comment below: > Section 1.2 paragraph 2: The matching of those operations a is ruled by > the order.... This is a bit confusing, as the paragraph mentions both > point-to-point operations and blocking collective operations. I would > suggest changing this to "The matching of those blocking collective > operations is rules by the order ...." fixed > Section 2: "High-quality ..." - In general I am against these sorts of > comments, even though they are strewn throughout the standard. There are > many tradeoffs in implementing a communications library, and what may be > good in one instance, may not be appropriate in another. It may be more > appropriate to state "The enables the application to take advantage of > asynchronous progress, if the implementation implements such a > capability..." While I agree with the sentiment, I disagree with the > categorization. yes/no: I'm hesitant to change this before we actually clarify the intent of those statements in the remaining standard. I will bring this up at the next meeting. > Section 2.1: Instead of using the term nested collectives, multiple > outstanding non-blocking collectives seems clearer to me. fixed > The comment that calling MPI_Request_free() is not > useful on the send side is not quite clear to me. Why is the send side > different from the receive side ? In either case, I do not think that we > should allow freeing a request in the middle of a collective (like one can > for ptp communications) The text actually states: "Freeing a request is only useful at the sender side and not on the receiver side." What I meant refers to MPI-2.1 page 55 lines 22-23: An active receive request should never be freed as the receiver will have no way to verify that the receive has completed and the receive buffer can be reused." I think request_free makes so many problems and is really not very useful. We should think again about deprecating it. But we should certainly not allow it for collectives! > The paragraph on the bottom of page 3 is confusing. After mentioning > that we can have multiple outstanding collectives, the last sentence seems > to imply that in such a case, all would have to be of the same type (such > as ibcast), which I do not think is the intent. this will be completely rephrased and improved with the examples from the last meeting in the new version. > Section 2.3: As I mentioned before, I do not believe we should specify > that an implementation should support more than a minimum of 1 (i.e must > provide support for this). Especially as the system sizes increase > markedly, we need to be careful on what sort of resource requirements we > place on an implementation. Hmm, I tend to disagree because if we don't guarantee anything, the it's really hard to write portable programs (i.e., all codes I know that use NBC would not be portable). Also, as we define it right now, it would be really easy to fall back to a software implementation which is available anytime. I would also claim that a software installation can be built with resource requirements that scale with O(1) with the number of processors (this implementation would be slow which is certainly bettar than not working at all). The interesting question is, how MPI-2.1 handles this issue with nonblocking point-to-point operations. Phrased differently: is an implementation that fails for every Isend/Irecv with "out of resources" a legal MPI-2.1 implementation? If yes, I would agree to the rephrasing, if no, then I would like to reconsider this topic. I know that we said in the Forum that this is outside the scope of MPI. But I think we should define a minimal quality (as we do for the tag space) in order to enable portable programming. Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

[mpi3-coll] Our MPI-2.2 Items
by Torsten Hoefler 09 Nov '08

09 Nov '08

Hello workgroup, our WG has several open MPI-2.2 items. I am listing them here for discussion. We should make progress on all of them (i.e., bring them to e review-ready state or request more feedback from the forum). We should discuss the remaining open items in the mailinglist and/or during our next teleconference (the Thursday after SC08). Here are the items: #23 (Generalized v-collectives: allow unknown recvcounts / rdispls): - seems to have several problems and the benefits seem to be very limited (main concerns: benefits only exist for small messages and an upper bound to the cumulative message size has to be known before the operation. So the user could as well do a non-vector operations with slightly larger buffers and achieve higher performance with negligible memory overhead. It is also unclear how to implement this performing (e.g., with a tree algorithm) without excessive memory copies and forcing serial receives.) #24 (Add a local Reduction Function): - seems to be ready for first reading - Jesper volunteered to add required changes to document, I will review #27 (Regular (non-vector) version of MPI_Reduce_scatter): - is missing a use-case, implementation and a clear benefit #30 (Clarification to intercomm MPI_Barrier): - awaiting reviews #31 (Add MPI_IN_PLACE option to Alltoall): - it's still unclear whether we want MPI_IN_PLACE or MPI_Alltoall_replace() - we should have a straw-vote during the next telecon (after SC) and update the proposal #33 (Fix Scalability Issues in Graph Topology Interface): - issue if we want weights (and what weights) or not is open - otherwise it seems ready for a first reading - we should come to a conclusion with the weight issues before proceeding and update the proposal #38 (Matched probe/receive): - seems to fit better in p2p WG - will move over #44 (Non blocking versus non-blocking versus nonblocking): - trivial, ready for reviews by chapter authors #26 (Add a callback function if a request completes): - relatively clear "no" for 2.2 - will move to p2p WG New additions are still welcome! All the best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608

1 0

Re: [mpi3-coll] Non-blocking Collectives Proposal Draft
by Richard Graham 19 Oct '08

19 Oct '08

Here are my comments: These are from the original draft sent out, so other may have already commented on some of these could not download from the web on the plane :-) Section 1.2 paragraph 2: The matching of those operations a is ruled by the order.... This is a bit confusing, as the paragraph mentions both point-to-point operations and blocking collective operations. I would suggest changing this to ³The matching of those blocking collective operations is rules by the order ....² Section 2: ³High-quality ...² - In general I am against these sorts of comments, even though they are strewn throughout the standard. There are many tradeoffs in implementing a communications library, and what may be good in one instance, may not be appropriate in another. It may be more appropriate to state ³The enables the application to take advantage of asynchronous progress, if the implementation implements such a capability...² While I agree with the sentiment, I disagree with the categorization. Section 2.1: Instead of using the term nested collectives, multiple outstanding non-blocking collectives seems clearer to me. The comment that calling MPI_Request_free() is not useful on the send side is not quite clear to me. Why is the send side different from the receive side ? In either case, I do not think that we should allow freeing a request in the middle of a collective (like one can for ptp communications) The paragraph on the bottom of page 3 is confusing. After mentioning that we can have multiple outstanding collectives, the last sentence seems to imply that in such a case, all would have to be of the same type (such as ibcast), which I do not think is the intent. Section 2.3: As I mentioned before, I do not believe we should specify that an implementation should support more than a minimum of 1 (i.e must provide support for this). Especially as the system sizes increase markedly, we need to be careful on what sort of resource requirements we place on an implementation. Rich On 10/16/08 4:46 PM, "Torsten Hoefler" <htor(a)cs.indiana.edu> wrote: > Hello Alexander, >> > Generally, some sentences start with lower case letters that should be >> > capitalized. > yes, I fixed some but will re-read it later this week. > >> > Page 3, top. We probably should not explain here how to ensure progress >> > for nonblocking ops. This matter should be covered by the general >> > progress description elsewhere. > I agree, I deleted it (it's already in the nonblocking point-to-point > chapter). > >> > Page 3, lower half. Look how progress is defined in this list by >> > referring to the pt2pt progress rules. > what do you mean with that? I would just reference the current MPI > definition - NBCs seem to fit this model nicely. > >> > Ibid. There are already 3 classes of requests: pt2pt, generalized, and >> > file I/O ones. > yes > >> > Page 3, bottom. How many nonblocking ops can overlap with a blocking >> > one? I.e., can one have the following situation: >> > >> > Process 0 Process 1 >> > >> > MPI_Ibarrier(req) MPI_Ibarrier(req) >> > MPI_Wait(req) >> > MPI Bcast MPI_Bcast >> > MPI_Wait(req) > yes, this is a legal code and should work. Any objections? > >> > Page 5, bottom. 32767 simultaneous ops to support is a tall order. I'd >> > say, 1 (one) would be a good lower limit, thus allowing no overlap. Or >> > even 0 (zero), meaning no nonblocking collective support at all. The >> > rest of this passage would probably be better reformulated as an advice >> > to implementors. > yes it is - I'm very indicisive. We have to discuss this at the Forum in > Chicago. > >> > In the margin: editing a PDF requires creativity (exemplified by the >> > earlier comments) or a pretty expensive Acrobat. Maybe we can find some >> > other way of distributing and commenting on our drafts? > oh yes, I fully agree. But the final document should be in MPI style, so > I thought I should start with LaTeX. Do you have any suggestions? Wiki > seems suboptimal because we would have to re-format everything (which > might be not too bad). LaTeX source in svn would be the best, but does > everybody have LaTeX? I'm open for any suggestion! > > For now, I uploaded the changed version to the wikipage (to avoid > attachments on the mailinglist): > https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/NBColl > > Thanks for the comments! > > Best, > Torsten > > -- > bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- > Torsten Hoefler | Postdoctoral Researcher > Open Systems Lab | Indiana University > 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA > Lindley Hall Room 135 | +01 (812) 855-3608 > _______________________________________________ > mpi3-coll mailing list > mpi3-coll(a)lists.mpi-forum.org > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll >

1 0

Re: [mpi3-coll] Non-blocking Collectives Proposal Draft
by Supalov, Alexander 17 Oct '08

17 Oct '08

Dear Rich, Thank you. I'm basically after making sure that we don't design something that does not work. From this discussion, it appears that the existing progress rules do cover the proposal well enough to proceed. Best regards. Alexander ________________________________ From: mpi3-coll-bounces(a)lists.mpi-forum.org [mailto:[email protected]] On Behalf Of Richard Graham Sent: Friday, October 17, 2008 3:42 PM To: MPI-3 Collective Subgroup Discussions Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft Alexander, I am not sure what the really trying to get after. The example I gave is exactly what I intended for it to be, not a complete working MPI example, but the core as it relates to progress. The example that Torsten gave should progress and complete, whether one uses asynchronous progress mechanisms or the synchronous approach most of us take today. If the question is about the semantics of MPI_Ibarrier(), I think the only specification is that it starts when the ibarrier is posted (what ever that means, which I am sure is implementation specific) and ends with the wait - I don't think that there is any other specification here. Have I completely missed what your question is ? Rich On 10/17/08 8:51 AM, "Supalov, Alexander" <alexander.supalov(a)intel.com> wrote: Dear Rich, Thanks. You probably meant Recv(0) in Proc 1. I also miss a Wait for req_r. So, a complete example would probably look like: Proc 0 Proc 1 Irecv(1,req_r) Irecv(0,req_r) Isend(1,req_s) Isend(0,req_s) Waitall(req_s, req_r) Recv(0) Send(1) Waitall(req_s,req_r) This would work. However, let's do one change, replacing Isend by Issend. Will this still work? Irecv(1,req_r) Irecv(0,req_r) Issend(1,req_s) Issend(0,req_s) Waitall(req_s, req_r) Recv(0) Send(1) Waitall(req_s,req_r) Another possible complication: let's swap the Send and Recv, and make Send synchronous (or rendezvous): Irecv(1,req_r) Irecv(0,req_r) Issend(1,req_s) Issend(0,req_s) Waitall(req_s, req_r) Ssend(0) Recv(1) Waitall(req_s,req_r) Here I have rather strong doubts that this will work. Do you? Best regards. Alexander -----Original Message----- From: mpi3-coll-bounces(a)lists.mpi-forum.org [mailto:[email protected]] On Behalf Of Graham, Richard L. Sent: Friday, October 17, 2008 2:35 PM To: mpi3-coll(a)lists.mpi-forum.org Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft Alexander, Do you see this as different in principal from Proc 0. Proc 1 irecv(1,req_r) Irecv(0,req_r) Isend(1,req_s). Isend(0,req_s) Wait(req_s). Send(0) Send(1). Wait(req_s) I see these two cases as basically the same from a progress perspective. Rich ----- Original Message ----- From: mpi3-coll-bounces(a)lists.mpi-forum.org <mpi3-coll-bounces(a)lists.mpi-forum.org> To: MPI-3 Collective Subgroup Discussions <mpi3-coll(a)lists.mpi-forum.org> Sent: Fri Oct 17 07:22:02 2008 Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft Dear Rich, Thank you. Note that the first operation is an Ibarrier. According to Torsten's explanation, process 0 will block in the Wait, which it may well do if it is Wait in process 1 that actually completes its Ibarrier. Now, how can process 0 make progress in its Wait when there's no matching Wait on the other side yet? Do we assume that any messages that may have been sent by process 0's IBarrier will be handled as unexpected messages by process 1 that is busy doing its (also unmatched) part of the Bcast? Likewise, do we assume that all messages that process 1 has to send in its Ibarrier can leave process 1 before it enters its Wait? Aren't we assuming a bit too much? In other words, aren't we assuming that all message sends and receives will be started in Ibarrier, and will be able to proceed even though the Wait is some blocking collective ops away? Given the way most MPIs implement nonblocking pt2pt ops now, something looks fishy here. In any case, I'd like to see a detailed explanation how things will work, and moreover, a proof that this does not preclude some algorithms - both in the sense of the pt2pt operations deployed, and in the sense of collective exchange pattern, for the applicable message sizes - from being used. Best regards. Alexander -----Original Message----- From: mpi3-coll-bounces(a)lists.mpi-forum.org [mailto:[email protected]] On Behalf Of Graham, Richard L. Sent: Friday, October 17, 2008 11:57 AM To: mpi3-coll(a)lists.mpi-forum.org Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft To me this does not look like a deadlock, proc o will make progress in the wait, and just like with point-to-point commuications process 1 will make general mpi progress while in the bcast. This includes progress on the ibcast so that process 0 can comete it's ibcast and move on to the bcast. Rich ----- Original Message ----- From: mpi3-coll-bounces(a)lists.mpi-forum.org <mpi3-coll-bounces(a)lists.mpi-forum.org> To: MPI-3 Collective Subgroup Discussions <mpi3-coll(a)lists.mpi-forum.org> Sent: Fri Oct 17 05:14:04 2008 Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft Dear Torsten, Thanks. I'm still unsure I understand how the example would work if processes blocked in the Wait calls as you said below. In this case process 0 would have no chance to reach the Bcast, while process 1 would be inside it and never reached the Wait. Here again is the original diagram: Process 0 Process 1 MPI_Ibarrier(req) MPI_Ibarrier(req) MPI_Wait(req)* MPI Bcast MPI_Bcast* MPI_Wait(req) The stars mark the point where the respective processes would block, as far as I understand. Looks like deadlock? Apart from this, stating that the nonblocking collectives can block in Wait presumes a certain implementation. What if Ibarrier was able to do all right away, and Wait were just a no-op? I think mixing blocking and nonblocking collectives in the proposed way should either be very carefully considered, or simply disallowed. Best regards. Alexander -----Original Message----- From: mpi3-coll-bounces(a)lists.mpi-forum.org [mailto:[email protected]] On Behalf Of Torsten Hoefler Sent: Thursday, October 16, 2008 11:15 PM To: MPI-3 Collective Subgroup Discussions Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft Hello Alexander, > Thank you. The pt2pt progress comment below was taken into account when > you removed the nonblocking collective progress description. ah, ok - so I think a reference to the chapter should be sufficient. > As for the example, if it is legal, we should probably count the > blocking operations as well when we decide how many simultaneous ops are > to be supported. yes, maybe. I am not sure but this could be an issue. It's on the agenda https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/forum201008 > Why "if" above: I'm not sure what a delayed nonblocking barrier would > mean for process 1. Won't one process possibly block somewhere? Can you > explain this please? the barriers will only block in the respective MPI_Wait() (or a tight MPI_Test() loop). But the operation can be initiated as soon as the critical region is left (and there mihgt be some other independent computation to do). > The format matter is complicated. I think it is worth discussing at the > Forum as we're likely to have more and more proposals approaching the > final stage, when editing them in PDF format will become an issue. yes, it's on the WG agenda - maybe we should discuss this in the whole forum? Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608 _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll --------------------------------------------------------------------- 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. _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll --------------------------------------------------------------------- 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. _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll --------------------------------------------------------------------- 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. _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll --------------------------------------------------------------------- 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.

1 0

Re: [mpi3-coll] Non-blocking Collectives Proposal Draft
by Supalov, Alexander 17 Oct '08

17 Oct '08

Got it, thanks. -----Original Message----- From: mpi3-coll-bounces(a)lists.mpi-forum.org [mailto:[email protected]] On Behalf Of Torsten Hoefler Sent: Friday, October 17, 2008 3:48 PM To: MPI-3 Collective Subgroup Discussions Subject: Re: [mpi3-coll] Non-blocking Collectives Proposal Draft Hello, > Thanks. You probably meant Recv(0) in Proc 1. I also miss a Wait for > req_r. So, a complete example would probably look like: > > Proc 0 Proc 1 > > Irecv(1,req_r) Irecv(0,req_r) > Isend(1,req_s) Isend(0,req_s) > Waitall(req_s, req_r) Recv(0) > Send(1) Waitall(req_s,req_r) > > This would work. However, let's do one change, replacing Isend by > Issend. Will this still work? > > Irecv(1,req_r) Irecv(0,req_r) > Issend(1,req_s) Issend(0,req_s) > Waitall(req_s, req_r) Recv(0) > Send(1) Waitall(req_s,req_r) sorry for joining in so late - it was night here. So I would still say that it doesn't deadlock - in the send/recv as well as in the collective case. Actually, from a correctness perspective, I see no difference between Rich's Isend() and your Issend() example because an MPI implementation is free to implement Isend() with Issend() (rendezvous protocol). Thus, implementers have to assume the worst case. To your example: - the process 1 is required to progress *all* outstanding operations while it is in Recv(0) - this enables the Waitall(req_s, req_r) on proc 0 to complete and proc 0 postes the send This is required by the MPI-2.1 standard afaik. > Another possible complication: let's swap the Send and Recv, and make > Send synchronous (or rendezvous): > > Irecv(1,req_r) Irecv(0,req_r) > Issend(1,req_s) Issend(0,req_s) > Waitall(req_s, req_r) Ssend(0) > Recv(1) Waitall(req_s,req_r) > > Here I have rather strong doubts that this will work. Do you? It's the same thing, Ssend will also have to progress the other requests. Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Researcher Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608 _______________________________________________ mpi3-coll mailing list mpi3-coll(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-coll --------------------------------------------------------------------- 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.

1 0