UCX F2F Meeting August 2015 - openucx/ucx GitHub Wiki

UCX face to face meeting, August 2015

When

Core discussions - August 11-13, 2015

Where

University of Knoxville TN, ICL Lab.
Philander P. Claxton Education Bldg., room 233
Knoxville, TN

Parking accommodations will be provided, the closest available visitor parking is G15 on White Avenue (see the following map, select "visitor" in the parking section: http://www.utk.edu/maps/ ).
Please use the parking visitor entrance on White Avenue and not the entrance on Clinch Ave (which is reserved for staff/students). Also please note that White Avenue is one way (you can use Clinch and then 16th South to get back to White). Park in this parking, we will have exit passes for you at the meeting.
Then go to Philander P Claxton.

Why: Agenda

Day 1 - Integration with advanced/emerging technologies.
09:00 AM

  • Presenting current UCX API (UCT,UCP) and discussing design decisions
  • Presenting PAMI
  • Endpoint scalability - reducing memory usage
  • Pending queue implementation
  • Multi-rail

Day 2 - Solving well-known problems in a better way
09:00 AM

  • Presenting Portals
  • Define thread safety model
  • Asynchronous progress and timers
  • Memory types, "routing" and GPU memory
  • Integration with Parsec, IO, MPICH

Day 3
09:00 AM

  • UCP blocking/nonblocking
  • UCP Active messages
  • Support for NVRAM
  • Spec - doxygen
  • Shared memory on different platforms
  • Summary and work plan for upcoming year
  • Registration cache structure

Laundry list

  • Collective APIs
  • Checkpoint/restart
  • Reducing progress overhead - avoid polling on unnecessary things

Organizers:

  • Pavel Shamis (ORNL)
  • Yossi Itigin (Mellanox)
  • George Bosilca (UTK)
  • Aurelien Bouteiller (UTK)

Attendees:

  • Pavel Shamis (ORNL)
  • Oscar Hernandez (ORNL)
  • Aurelien Bouteiller (UTK)
  • Yossi Itigin (Mellanox)
  • Sameh Sharkawi (IBM)
  • Graham Lopez (ORNL)
  • Sreeram Potluri (NVIDIA)
  • George Bosilca (UTK)
  • Manjunath Gorentla (ORNL)
  • Amith Mamidala (IBM)
  • Sameer (IBM)
  • Matthew Baker (ORNL)
  • Stephen Richmond (UTK)
  • Alina Sklarevich (Mellanox)
  • Lena Oden (ANL)
  • Edgar Leon (LLNL)
  • Kento Sato (LLNL)
  • Nicholas Park (DoD)
  • Ron Brightwell (Sandia)
  • Hasan Abbasi (ORNL)
  • Neena (ORNL)

Notes

  • Statist aggregation interface for PEs
  • Ordering semantics
  • UCP/UCT Do we want to share the context for memory caches or not ? Shell it happen automatically or we may provide knob to user
  • Endpoint:
  • Requirement - very fast endpoint creation and destroy.
  • Endpoint scalability (memory footprint)
  • MPI-4 Endpoints ? What to do about this ?
  • Constant key (predefined set of keys stored in a predefined location)
  • bcopy - better name ?
  • pending queue ?
  • do we need blocking interface in UCP ?
  • datatypes ?

Endpoint

  • UCP endpoint structure has to be compressed … and a lot. Trying to come up with a solution that can fit into 64bit. tbd