DARPA Data Reorganization Standards Effort Minutes

January 13, 1998
Held at Sandia National Labs, Albuquerque, NM
Local Host: Ron Brightwell

Attendance List:

NAMEORGANIZATIONEMAIL ADDRESS
Joe FoglerUNM/AHPCCfogler@arc.unm.edu
Steve PaavolaSky ComputersPaavola@sky.com
Randall JuddSPAWARSYSCENjudd@nosc.mil
Dennis CottelSPAWARSYSCENdennis@spawar.navy.mil
Thomas RobeyKhoral Researchtrobey@khoral.com
Ken CainMITREkcain@mitre.org
Jon GreeneMercurygreene@mc.com
Clay Taylor, Jr.MPI Soft. Tech.cdtaylor@mpi-softtech.com
Ron BrightwellSandiabright@cs.sandia.gov
Anna RounbehlerRaytheon Electronic SystemsAnna_C_Roubehler@res.raytheon.com

Minutes:

The meeting came to order at 9:00am

Taylor

Introduction, the agenda for the day.

Rounbehler

Presentation, "Optimizing Popular Classes of Algorithms"

[Not yet available on the web page]

Dennis, Randal

What is the definition of API as related to this effort?

Anna

The effort may move toward defining a standard, based on the committee's decision. See charter of this effort on the home page.

Steve

Lincoln Labs is working on an API for parallel data reorganization.

Dennis

Several API's already exist, we need to describe data some way.

Jon, Randal, Joseph

(from Anna's presentation) We need to define domain. Vector domain matrix domain, tensor domain, and also consider methods of communication that are required.

Randal, Anna, Dennis

The purpose of this group should be for new API's or those that are still under development.

Kenneth

We must consider data distribution, some methods could become a foundation for more complicated methods.

Dennis, Randal

There are some data conversion issues, different standards and different algorithms for an individual task will require opposing data organizations.

Jon

There really is not a need for the more esoteric reorganization methods, just consider the more popular data organization requirements.

Jon, Randal

Data may be described one way on one side, but may need to be accessed in a different way on the other side. We need an API that will marry the 2 distribution requirements into one data object.

break

Kenneth

Example of a benchmark for a distributed array, operation to be done "in-place", that is in the same process group.

Jon

Example of a corner-turn operation, issues are divisibility and alignment. Sometimes big contiguous move followed by more complex reorganization is best.

Lunch break

Taylor

Opened afternoon discussion.

Paavola

Layout of data object and streams reorganization model.

Dennis

Layout of random access model

Some important reorganization problems mentioned were cases where the sizes change, such as the MPEG format. The streams concept is important for cases where the required data size is too large to fit into memory.
Thom, then others
Layout of documentation framework.

Please forward any questions/comments/corrections/suggestions to
cdtaylor@mpi-softtech.com (Clayborne Taylor, Jr.)