Thanks. You are right - if there's more than one route between two processes, there's a matching issue, too. As for my special implementor's point of view, I was kidding. ________________________________ From: mpi3-subsetting-bounces_at_[hidden] [mailto:mpi3-subsetting-bounces_at_[hidden]] On Behalf Of Richard Graham Sent: Friday, February 29, 2008 4:20 PM To: mpi3-subsetting_at_[hidden] Subject: Re: [Mpi3-subsetting] Some "stupid user" questions, comments. On 2/29/08 9:26 AM, "Supalov, Alexander" <alexander.supalov_at_[hidden]> wrote: Dear RIchard, Thanks. The more complicated the standard gets, the happier are the implementors. However, now we try to think like MPI users for a change, so, thanks for providing a reality check. >> Quite to the contrary. The simpler the standard is the easier to support - complexity is not a good thing at all. >> This is my view as an implementer. Complexity is often introduced when trying to get good performance out of >> a spec that supports a wide variety of options. Now, to one of your questions. An MPI_ANY_SOURCE MPI_Recv in multifabric environment means that a receive has to be posted somehow to more than one fabric in the MPI device layer. Once one of them gets the message, the posted receives should be cancelled on other fabrics. Now, what if they've already matched and started to receive something? What if they cannot cancel a posted receive? And so on. There are 3 to 5 ways to deal with this situation, with and without actually posting a receive, but none of them is good enough if you ask me. That's why there are 3 to 5 of them, actually. And all of them complicate the progress engine - the heart of an MPI implementation - at exactly the spot where one wants things simple and fast. >> The any_source and multiple fabrics are two distinct issues. Even if you do not support any_source and have >> multiple fabrics, you have the issue that to support mpi ordering semantics, matching needs to be done >> in the context of all the nics - unless you decide to have only one nic do the matching, including any on-host >> traffic. What any_source forces is matching on the receive side - unless one wants to set up a very complex >> and inefficient way to make sure that only one receive is matched for each wild card receive. Rich This means that most of the time we fight these repercussions and curse the MPI_ANY_SOURCE. Or, looping back to the beginning of this message, we actually never stop blessing MPI_ANY_SOURCE. Fighting this kind of trouble is what we are paid for. ;) Best regards. Alexander ________________________________ From: mpi3-subsetting-bounces_at_[hidden] [mailto:mpi3-subsetting-bounces_at_[hidden]] On Behalf Of Richard Barrett Sent: Friday, February 29, 2008 2:50 PM To: mpi3-subsetting_at_[hidden] Subject: [Mpi3-subsetting] Some "stupid user" questions, comments. Hi folks, I'm still sorting things out in my mind, so perhaps this note is just me talking to myself. But should you feel so compelled to sort through it, I would appreciate any feedback you might offer; and it will make me a more informed participant. I see two main perspectives: the user and the implementer. I come from the user side, so I feel comfortable in positing that user confusion over the size of the standard is really a function of presentation. That is, most of us get our information regarding using MPI directly from the standard. For me, this is the _only_ standard I've ever actually read! Perhaps I am missing out on thousands of C and Fortran capabilities, but sometimes ignorance is bliss. That speaks highly to the MPI specification presentation; however it need not be the case. An easy solution to the "too many routines" complaint is a tutorial/book/chapter on the basics, with pointers to further information. And in fact these books exist. That said, I hope that MPI-3 deprecates a meaningful volume of functionality. >From the implementer perspective, there appear to be two goals. First is to ease the burden with regard to the amount of functionality that must be supported. (And we users don't want to hear of your whining, esp. from a company the size of Intel :) Second, which overlaps with user concerns, is performance. That is, by defining a small subset of functionality, strong performance (in some sense, e.g. speed or memory requirements) can be realized. At the risk of starting too detailed a discussion at this early point (as well as exposing my ignorance:), I will throw out a few situations for discussion. 1. What would such a subset would imply with regard to what I view as support functionality, such as user-defined datatypes, topologies, etc? Ie could this support be easily provided, say by cutting-and-pasting from the full implementation you will still provide? (I now see Torsten recommends excluding datatypes, but what of other stuff?) 2. Even more broadly (and perhaps very ignorantly), can I simply link in both libraries, like -lmpi_subset -lmpi, getting the good stuff from the former and the excluded functionality from the latter? In addition to the application developers use of MPI, all large application programs I've dealt with make some use of externally produced libraries (a "very good thing" imo), which probably exceed the functionality in a "subset" implementation. 3. I (basically) understand the adverse performance effects of allowing promiscuous receives (MPI_ANY_SOURCE). However, this is a powerful capability for many codes, and used only in moderation, eg for setting up communication requirements (such as communication partners in unstructured, semi-structured, and dynamic mesh computations). In this case the sender knows its partner, but the receiver does not. A reduction(sum) is used to let each process know the number of communication partners from which it will receive data, the process posts that many promiscuous receives, which when satisfied lets it from then on specify the sender. So would it be possible to include this capability in a separate function, say the blocking send/recv, but not allow it in the non-blocking version? 4. Collectives: I can't name a code I've ever worked with that doesn't require MPI_Allreduce (though I wouldn't be surprised to hear of many), and this in a broad set of science areas. MPI_Bcast is also often used (but quite often only in the setup phase). I see MPI_Reduce used most often to collect timing information, so MPI_Allreduce would probably be fine as well. MPI_Gather is often quite useful, as is MPI_Scatter, but again often in setup. (Though often "setup" occurs once per time step.) Non-constant size versions are often used. And others can also no doubt offer strong opinions regarding inclusion of exclusion. But from an implementation perspective, what are the issues? In particular, is the basic infrastructure for these (and other collective operations) the same? A driving premise for supporting collectives is that the sort of performance driven capability under discussion is most needed by applications running at very large scale, which is where even very good collect implementations run into problems. 5. Language bindings and perhaps other things: With the expectation/hope that full implementations continue to be available, I could use them for code development, thus making use of things like type checking, etc. And does this latter use then imply the need for "stubs" for things like the (vaporous) Fortran bindings module, communicators (if only MPI_COMM_WORLD is supported), etc.? And presuming the answer to #2 is "no", could/should the full implementation "warn" me (preferably at compile time) when I'm using functionality that rules out use of the subset? 6. Will the profile layer still be supported? Generating usage can still be quantified using a full implementation, but performance would not be (at least in this manner), which would rule out an apples-to-apples comparison between a full implementation and the subset version with its advertised superior performance. (Of course an overall runtime could be compared, which is the final word, but a more detailed analysis is often preferred.) 7. If blocking and non-blocking are required of the subset, aren't these blocking semantics? MPI_Send: MPI_Isend ( ..., &req ); MPI_Wait ( ..., &req ); ----- MPI_Recv: MPI_Irecv ( ..., &req ); MPI_Wait ( &req ); - And speaking of this, are there performance issues associated with variants of MPI_Wait, eg MPI_Waitany, MPI_Waitsome? Finally, I'll officially register my concern with what I see as an increasing complexity in this effort, esp wrt "multiple subsets". I don't intend this comment to suppress ideas, but to help keep the beating the drum for simplicity, which I see as a key goal of this effort. If you read this far, thanks! My apologies if some of these issues have been previously covered. And if I've simply exposed myself as ignorant, I feel confident is stating that I am not alone - these questions will persist from others. :) Richard --------------------------------------------------------------------- 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. *