February 2009 - mpi-forum - mailman3.cels.anl.gov

Re: [Mpi-forum] Persistence Working Group Proposed
by Anthony Skjellum 25 Feb '09

25 Feb '09

Dick, there are plenty of one-end optimizations that are possible though. The current technology is "half channels" and we want "whole channels" too. Binding an actual channel (which I did propose in MPI-2) is also possible for further optimization, and I am open to that as well. That makes a FIFO between sender and receiver, with single credit. I agree that a channel is desirable also. We would make it part of the discussion. It is easy enough to specify... Tony ----- Original Message ----- From: "Richard Treumann" <treumann(a)us.ibm.com> To: "Main MPI Forum mailing list" <mpi-forum(a)lists.mpi-forum.org> Sent: Wednesday, February 25, 2009 12:11:30 PM GMT -06:00 US/Canada Central Subject: Re: [Mpi-forum] Persistence Working Group Proposed Tony In my opinion the lack of enthusiasm for point to point persistent operations does not tell us much about the usefulness of persistent operations in general. When the MPI-1 Forum decided an MPI_Isend_init or MPI_Irecv_init must be a local operation and that an MPI_Start of a persistent send on one side must match any form of receive (persistent or not) on the other (and vice versa) most of the opportunity for optimization was lost. For a persistent collective to be fully optimized, the MPI library probably must be able to do collective communication at the CC_INIT and return to each task a persistent request handle that is irrevocably bound to all the handles created at the other tasks by this collective operation. It should not be valid to create two persistent BCAST requests at the root with different send buffers and one persistent BCAST at the non-roots just so root could BCAST from either buffer. (even if somebody thinks that would be a cool bit of generality) 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-forum mailing list mpi-forum(a)lists.mpi-forum.org http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum -- Anthony Skjellum, PhD Professor and Chair Dept. of Computer and Information Sciences University of Alabama at Birmingham +1-(205)934-8657; FAX: +1- (205)934-5473 ___________________________________________ CONFIDENTIALITY: This e-mail and any attachments are confidential and may be privileged. If you are not a named recipient, please notify the sender immediately and do not disclose the contents to another person, use it for any purpose or store or copy the information in any medium. Please consider the environment before printing this e-mail

1 0

Re: [Mpi-forum] Persistence Working Group Proposed
by Richard Treumann 25 Feb '09

25 Feb '09

Tony In my opinion the lack of enthusiasm for point to point persistent operations does not tell us much about the usefulness of persistent operations in general. When the MPI-1 Forum decided an MPI_Isend_init or MPI_Irecv_init must be a local operation and that an MPI_Start of a persistent send on one side must match any form of receive (persistent or not) on the other (and vice versa) most of the opportunity for optimization was lost. For a persistent collective to be fully optimized, the MPI library probably must be able to do collective communication at the CC_INIT and return to each task a persistent request handle that is irrevocably bound to all the handles created at the other tasks by this collective operation. It should not be valid to create two persistent BCAST requests at the root with different send buffers and one persistent BCAST at the non-roots just so root could BCAST from either buffer. (even if somebody thinks that would be a cool bit of generality) 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

1 0

Persistence Working Group Proposed
by Anthony Skjellum 25 Feb '09

25 Feb '09

Everyone, I would like to reintroduce proposals for a) persistent collectives (including persistent non-blocking) b) persistent generalized requests c) ability to modify persistent operations d) techniques to allow "initialization" vs. "instance" performance tradeoffs to be expressed portably All of these things help with performance and scalability, and do little to implementation complexity where they are not done. Persistent operations are very useful with data parallel and pipelined parallel programs with temporal locality and uniform or slowly changing patterns of communication. Static and semi-static patterns are common and can be further optimized. Persistence takes "middle level code" out of the critical path, and encourages concurrent operations, when used in combination with non-blocking point to point and collective technology. Persistence allows planned-transfer and deep optimizations to be done "one time" so that the instances are faster. This has proven very successful with technologies such as FFTW, and is common to low-level message passing systems used in some embedded HPC machines. Persistence allows for reducing or eliminating some of the costs of memory management as well as for choosing better collective algorithms . For example, optimized strategies for pinning, unpinning, and streamlining derived datatypes become a snap with persistent mode. Adding persistent collective (blocking and non-blocking) and persistent generalized requests orthogonalizes the design of MPI in a convenient way as well, but it is for performance and scalability that it is most important. While we already have send_init() and recv_init() functionality, it is limited to point-to-point. Commercial implementations have in past found scope for getting more performance out of persistent than point to point; early implementations chose not to do so. As architectures evolve, and memory management becomes more complex and hierarchical, including DMA architectures, and one wants to run MPI at higher and higher scales, planned transfer modes for regular/data-parallel programs will expand the performance potential of MPI programs, while requiring relatively simple transformations of existing regular parallel programs. I will finally mention that persistent generalized requests are useful so as to allow complex patterns to be reused once specified, rather than having to be reconstructed at each use. I will argue that persistent generalized requests are more useful than non-persistent ones in many many way. :-) I'd like to propose a working group for that. It fits squarely in MPI 3.0. Rolf observed that in past, few people use persistent operations in their studies. However, it may be because it had little performance benefit in then existing implementations. Moving forward, and when applied to collective operations, huge performance gains are possible, as they also are when derived datatypes are part of point-to-point. I hope other institutions will agree so we can pursue this. Regards, Tony -- Anthony Skjellum, PhD Professor and Chair Dept. of Computer and Information Sciences University of Alabama at Birmingham +1-(205)934-8657; FAX: +1- (205)934-5473 ___________________________________________ CONFIDENTIALITY: This e-mail and any attachments are confidential and may be privileged. If you are not a named recipient, please notify the sender immediately and do not disclose the contents to another person, use it for any purpose or store or copy the information in any medium. Please consider the environment before printing this e-mail

1 0

Re: [Mpi-forum] April Forum meeting
by Richard Graham 25 Feb '09

25 Feb '09

BTW, Please send me requests for working sessions. I will use last meeting's agenda as a starting point. Rich On 2/24/09 12:18 PM, "Jeff Squyres" <jsquyres(a)cisco.com> wrote: > Logistics for the April meeting are now available (e.g., address of > meeting site, registration for the meeting/hotel, etc.): > > http://meetings.mpi-forum.org/Apr_2009_logistics.php

1 0

Re: [Mpi-forum] April Forum meeting
by Richard Graham 25 Feb '09

25 Feb '09

The cutoff date for hotel reservation is March 16th, so please sign up by then. In the past hotels have been eager for business, and have kept taking reservations past the cutoff dates, and I expect this could be the case for this meeting also. Rich On 2/24/09 12:18 PM, "Jeff Squyres" <jsquyres(a)cisco.com> wrote: > Logistics for the April meeting are now available (e.g., address of > meeting site, registration for the meeting/hotel, etc.): > > http://meetings.mpi-forum.org/Apr_2009_logistics.php

1 0

April Forum meeting
by Jeff Squyres 24 Feb '09

24 Feb '09

Logistics for the April meeting are now available (e.g., address of meeting site, registration for the meeting/hotel, etc.): http://meetings.mpi-forum.org/Apr_2009_logistics.php -- Jeff Squyres Cisco Systems

1 0

Euro PVM/MPI'09 CFP
by Jeff Squyres 24 Feb '09

24 Feb '09

Apologies if you have seen this already / please spread the CFP around. Thanks! EuroPVM/MPI 2009 CALL FOR PAPERS 16th European PVMMPI Users' Group Meeting Espoo, Finland, September 7 - 10, 2009 Web: http://www.csc.fi/english/pages/pvmmpi09/ Organized by CSC - IT Center for Science Submission website: http://www.easychair.org/conferences/?conf=europvmmpi2009 BACKGROUND AND TOPICS PVM (Parallel Virtual Machine) and MPI (Message Passing Interface) have become the standard interfaces for high-performance parallel programming in the message-passing paradigm. EuroPVM/MPI is the most prominent meeting dedicated to the latest developments of PVM and MPI, their use, including support tools, and implementation, and to applications using these interfaces. The EuroPVM/MPI meeting also encourages discussions of new message-passing and other parallel and distributed programming paradigms beyond MPI and PVM. The 16th European PVM/MPI Users' Group Meeting will be a forum for users and developers of PVM, MPI, and other message-passing programming environments. Through the presentation of contributed papers, vendor presentations, poster presentations and invited talks, attendees will have the opportunity to share ideas and experiences to contribute to the improvement and furthering of message-passing and related parallel programming paradigms. Topics of interest for the meeting include, but are not limited to: * PVM and MPI implementation issues and improvements * Latest extensions to PVM and MPI * PVM and MPI for high-performance computing, clusters and grid environments * New message-passing and hybrid parallel programming paradigms, eg. with general purpose graphics processors * Interaction between message-passing software and hardware * Fault tolerance in message-passing programs * Performance evaluation of PVM and MPI applications * Tools and environments for PVM and MPI * Algorithms using the message-passing paradigm * Applications in science and engineering based on message-passing Submissions on applications demonstrating both the potential and shortcomings of MPI and PVM are particularly welcome. As in the previous years, the special session 'ParSim' will focus on numerical simulation for parallel engineering environments. EuroPVM/MPI 2009 will also hold the 'Outstanding Papers' session, where the best papers selected by the program committee will be presented. SUBMISSION INFORMATION Contributors are invited to submit a full paper as a PDF document not exceeding 10 pages in English (2 pages for poster abstracts). In addition, two more pages may be included on a page charge basis, 50 euros/page. The title page should contain an abstract of at most 100 words and five specific keywords. The paper needs to be formatted according to the Springer LNCS guidelines (http://www.springer.de/comp/lncs/authors.html) The usage of LaTeX for preparation of the contribution as well as the submission in camera ready format is strongly recommended. Style files can be found at http://www.springer.de/comp/lncs/authors.html. New work that is not yet mature for a full paper, short observations, and similar brief announcements are invited for the poster session. Contributions to the poster session should be submitted in the form of a two page abstract. All these contributions will be fully peer reviewed by the program committee. Papers are to be submitted electronically via the online submission system at http://www.easychair.org/conferences/?conf=europvmmpi2009. Submissions to the special session 'Current Trends in Numerical Simulation for Parallel Engineering Environments' (ParSim 2009) are handled and reviewed by the respective session chairs. For more information please refer to the ParSim website (http://www.lrr.in.tum.de/Par/arch/events/parsim09/) All accepted submissions are expected to be presented at the conference by one of the authors, which requires registration for the conference. IMPORTANT DATES EuroPVM/MPI Conference Submission of full papers and poster abstracts April 8th, 2009 Notification of authors May 11th, 2009 Camera-ready papers June 8th, 2009 Tutorials Monday, September 7th, 2009 Conference Tuesday-Thursday, September 8th-10th, 2009 For up-to-date information, visit the conference web site at http://www.csc.fi/english/pages/pvmmpi09/. ParSIM Session Submission of papers May 6th, 2009 Notification to Authors May 26th, 2009 Camera ready papers June 8th, 2009 PROCEEDINGS The conference proceedings consisting of abstracts of invited talks, full papers, and two page abstracts for the posters will be published by Springer in the LNCS series. Authors are strongly encouraged to read carefully the recommendations for publications to facilitate the publication procedure. General chair: Jack Dongarra (University of Tennessee) Programme Committee chair: Jan Westerholm (ï¿½bo Akademi University) PC co-chairs: Juha Fagerholm and Jussi Heikonen (CSC) Local organizing committee chair: Per ï¿½ster (CSC) CONFERENCE VENUE The conference is held the Swedish-Finnsih Cultural Centre in Hanasaari, Espoo, Finland (www.hanaholmen.fi) The conference is organized and hosted by CSC - IT Center for Science Ltd. -- Jeff Squyres Cisco Systems

1 0

wordsmiths wanted for MPI glossary
by Terry Jones 15 Feb '09

15 Feb '09

All, I am putting together a glossary to be incorporated into upcoming MPI specs (see ticket #78, https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/78#comment:3) Please take a moment to review the current draft (see attached) and send me your comments. To avoid spamming this list, PLEASE simply email comments directly to me (trj(a)ornl.gov) or place your comments on the wiki at item #78. All comments, suggestions, and corrections are welcome. Thank you. Regards, Terry trj(a)ornl.gov

1 0

Re: [Mpi-forum] New field on Trac tickets
by Jeff Squyres 14 Feb '09

14 Feb '09

On Feb 13, 2009, at 9:51 AM, Jeff Squyres wrote: > We may add another field or two shortly; these new fields will also > be required to be set manually. Hopefully we'll decide on the new > field(s) Very Soon so that you can update all the new fields at once. It looks like it will be easier to hold off on any additional fields until MPI-3. So "implementation" is the only new one for now. -- Jeff Squyres Cisco Systems

1 0

Re: [Mpi-forum] Votes posted
by Jeff Squyres 13 Feb '09

13 Feb '09

Oops -- I meant to post the general "Feb 2009 meeting" link: http://meetings.mpi-forum.org/secretary/2009/02/ On Feb 13, 2009, at 9:28 AM, Jeff Squyres wrote: > General notes and votes from the San Jose meeting have been posted. > I'll post slides when I get them. > > http://meetings.mpi-forum.org/secretary/2009/02/agenda.php -- Jeff Squyres Cisco Systems

1 0