Skip to content

Home

Jump to bottom
Guillaume Mercier edited this page Jan 9, 2023 · 91 revisions

Planned telcos/meetings

Next Telcos:

  • January the 13th 2022, 3pm (Paris time).
  • January the 27th 2022, 3pm (Paris time).

Recordings of past meetings

The recordings of past meetings are available on the wiki of the mpi-standard repository of the Working Group (the other repository of this WG). Since this repo is private you'll need to register in order to access it.

Past Telcos:

  • 2018/03/29 (Thursday): EU/JP telco at 9am/GMT+2
  • 2018/04/05 (Thursday): JP/US telco at 1am/GMT+2
  • 2018/04/12 (Thursday): EU/US telco at 5pm/GMT+2
  • 2018/04/26 (Thursday): EU/JP telco at 9am/GMT+2
  • 2018/05/03 (Thursday): US/JP telco at 1am/GMT+2 Cancelled
  • 2018/05/17 (Thursday): EU/US telco at 4pm/GMT+2
    • mpibind by Edgar Leon
  • 2018/05/31 (Thursday): EU/JP telco at 9am/GMT+2
    • netloc by Brice Goglin
  • 2018/06/07 (Thursday): US/JP telco at 1am/GMT+2 Cancelled
  • 2018/06/12: F2F meeting in Austin (9:00 am -> 12:30 am local time)
    • Fujitsu's network topology aware extension by Shinji Sumimoto
  • 2018/06/21 (Thursday): EU/JP telco at 9am/GMT+2 => postponed 2018/06/22 10am/GMT+2
  • 2018/06/28 (Thursady): US/JP telco at 1am/GMT+2
  • 2018/07/26 (Thursday): EU/JP telco at 9am/GMT+2
  • 2018/08/09 (Thursday): EU/US telco at 4pm/GMT+2
  • 2018/09/19 (Wednesday) : F2F meeting in Barcelona
  • 2018/10/10 (Wednesday): EU/JAP telco at 9am/GMT+2 Cancelled
  • 2018/10/24 (Wednesday): EU/US telco at 4pm/GMT+2
  • 2018/11/14 (Wednesday): EU/JAP telco at 9am/GMT+2
    • Moved to 4pm/GMT+1
  • 2018/11/28 (Wednesday): EU/US telco at 4pm/GMT+1
    • Moved to 4pm/GMT+1 Cancelled
  • 2018/12/04 (Tuesday) : F2F meeting in San José
  • 2018/12/19 (Wednesday): EU/JAP telco at 9am/GMT+1
  • 2019/01/09 (Wednesday): EU/US telco at 4pm/GMT+1 Cancelled
  • 2019/01/23 (Wednesday): EU/JAP telco at 9am/GMT+1
  • 2019/02/06 (Wednseday): EU/US telco at 4pm/GMT+1
  • 2019/02/20 (Wednesady): EU/JAP telco at 9am/GMT+1
  • 2019/03/27 (Wednesday): EU/USJAP telco at 4pm/GMT+1 **moved to 9:30am/GMT+1**
  • 2019/04/03 (Wednesday): MPI Forum Virtual Meeting at 5pm/GMT+2 about split_types
  • 2019/04/1011 (Wednesday Thursday): EU/JAP telco at 9am/GMT+2 **moved to 04/11 at 9am/GMT+2: Rolf's proposal (part one) **
  • 2019/04/15 (Monday): EU/JAP telco at 9am/GMT+2 -- **Rolf's proposal (part deux)**
  • 2019/04/24 (Wednesday): MPI Forum Virtual Meeting at 5pm/GMT+2 about cartesian topologies (issue #120)
  • 2019/05/03 (Friday): EU/US telco at 4pm/GMT+2 -- Rolf's proposal (part three)
  • 2019/05/09 (Thursday): EU/Jap Telco at 9am/GMT+2 -- Comm_split_type proposal **=> postposned to 2019/05/10 (Friday) at 9am/GMT+2**
  • 2019/05/29 (Wednesday): F2F meeting in Chicago at 1:30pm (Chicago time)
  • 2019/08/07 (Wednesday): EU/Jap Telco at 9am/GMT+2 =>postponed to Thursday 08/08 at 4pm/GMT+2
  • 2019/09/05-06 (Thursday or Friday): F2F meeting in Zürich (date and time TBD)
  • 2019/09/08 (Sunday): F2F meeting in Zürich at 9am/GMT+2
  • 2019/09/18 (Wednesday): EU/US telco at 4pm/GMT+2 => 3pm/GMT+2 *Preparation for the upcoming Virtual Meeting
  • 2019/09/25 (Wednesay): EU/Jap Telco at 9am/GMT+2 => 10am/GMT+2
    • Preparation for the upcoming Virtual Meeting
  • 2019/10/09 (Wednesday): EU/US telco at 4pm/GMT+2
    • Discussion about the query function MPI_Get_hw_subdomain_names (ticket #154)
  • 2019/10/16 (Wednesday): EU/Jap Telco at 10am/GMT+2 => Moved to 10:30am/GMT+2
    • Discussion about tickets #132 and #154, preparation for the 10/23 VM
  • 2019/10/21 (Monday): EU/Jap Telco at 10am/GMT+2
    • Discussion about tickets #132 and #154, preparation for the 10/23 VM (part deux)
  • 2019/11/06 (Wednesday): EU/US telco at 4pm/GMT+1
    • Preparation for the Albuquerque Physical Meeting
  • 2019/11/20 (Wednesday): EU/Jap Telco at 10am/GMT+1 => 4pm/GMT+1
    • Preparation for the Albuquerque Physical Meeting
  • 2020/02/18 (Tuesday): Physical Meeting at Portland 2pm (local time)
  • 2020/03/18 (Wednesday): EU/Jap Telco at 10am/GMT+1
  • 2020/04/01 (Wednesday): EU/US Telco at 5pm/GMT+2 => Postponed to Thursday, April the 2nd, 5pm/GMT+2
  • 2020/04/15 (Wednesday): EU/JAP at 10am/GMT+2

Hardware Topology Working Group Statement

As hardware is rapidly evolving, a wide spectrum of hardware features became available: complex memory hierarchies, accelerators, many-core processors, FPGAs, MPI on hardware, non-coherent memory, heterogeneous architectures, etc. As a consequence, understanding such features and exploiting them at the MPI application level (e.g. data locality) is a challenge that applications envision to address directly. Indeed, MPI still relies on (or enforces) a flat programming model, which is becoming less relevant with the advent of the heterogeneous many-core era. Despite recent modifications to the standard to address these scalability issues (e.g. distributed graph interface, neighborhood collectives, etc.), the challenges of hardware topology discovery and management have been left unaddressed by the MPI Forum so far.

The main question this Working Group proposes to address is the following: how can hardware resources (I/O, cores, caches, I/O proxies, etc.) be discovered, queried upon and distributed between execution flows? Indeed, MPI processes are bound to span multiple cores and no portable primitives exist outside, nor inside, MPI to explore and take advantage of the hardware topology either at the node or that is made available at the process level.

Assuming the need to expose such information to the application, here is a more detailed list of questions to be addressed in the working group:

  • What level of abstraction is desirable in order to keep the hardware-independence of MPI effective?
  • Which MPI objects/constructs could benefit and exploit such a knowledge (e.g groups, communicators, shared-memory windows, etc)?
  • How can we envision interactions between MPI and some other programming model/interface (e.g. OpenMP) with regard to hardware topology utilization?
  • How can this WG interact with other WGs that already started to address the same kind of issues, especially for shared memory?
  • How does this impact other discussed proposals, such as sessions and endpoints for instance?
  • More generally speaking, what kind of interactions are expected with the hybrid WG?
  • What should be the relationship between the proposed interface and the mapping/binding mechanisms provided by process managers and/or RJMS? Can we (should we?) standardize this also?

Despite such a wide range of questions, answering them is critical for the understanding of hierarchical notions at the application level. We plan to first address the hierarchical communicators interface proposal, as presented informally at the MPI Forum meeting in Portland (Feb 2017) by Inria, and from there open the discussion to the other topics mentioned above.

We envision open and explorative discussions around the hardware topology discovery and management idea as members will have the opportunity to provide their existing knowledge and feedback.

Tele-con Minutes

Clone this wiki locally