File: heterogeneity.rst

package info (click to toggle)
openmpi 5.0.8-4
  • links: PTS, VCS
  • area: main
  • in suites:
  • size: 201,684 kB
  • sloc: ansic: 613,078; makefile: 42,353; sh: 11,194; javascript: 9,244; f90: 7,052; java: 6,404; perl: 5,179; python: 1,859; lex: 740; fortran: 61; cpp: 20; tcl: 12
file content (21 lines) | stat: -rw-r--r-- 843 bytes parent folder | download | duplicates (8)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
.. _label-heterogeneous-mpi-apps:

Building heterogeneous MPI applications
=======================================

Heterogeneous support (specifically: supporting different sized and/or
represented data types in a single MPI application run) within a
single MPI job is technically required by the MPI standard.

However, there are both theoretical and practical problems with
supporting true data heterogeneity at run-time.

Indeed, it is quite uncommon for production HPC environments to be
data-heterogeneous (e.g., natively support little-endian on some nodes
and big-endian on other nodes in the same MPI application job).

As such, supporting data heterogeneity is a feature that has fallen
into disrepair and is currently known to be broken in this release of
Open MPI.

:doc:`Contributions to fix it would be welcome! </contributing>`