Changes in 4.1
-
Thread-cs in ch4 changed to per-vci.
-
Testsuite (test/mpi) is configured separately from mpich configure.
-
Added options in autogen to accelerate CI builds, including using pre-built
sub-modules. Added -yaksa-depth option to generate shallower yaksa pup code
for faster build and smaller binaries. -
Support singleton init using hydra.
-
On OSX, link option flat_namespace is no longer turned on by default.
-
Generate mpi.mod Fortran interfaces using Python 3. For many compilers,
including gfortran, flags such as -fallow-mismatched-args is no longer
necessary. -
Fixed message queue debugger interface in ch4.
-
PMI (src/pmi) is refactored as a subdir and can be separately distributed.
-
Added MPIX_Comm_get_failed.
-
Experimental MPIX stream API to enable explicit thread contexts.
-
Experimental MPIX gpu enqueue API. It currently only supports CUDA streams.
-
Delays GPU resource allocation in yaksa.
-
CH3 nemesis ofi netmod is removed.
-
New collective algorithms. All collective algorithms are listed in
src/mpi/coll/coll_algorithms.txt -
Removed hydra2. We will port unique features of hydra2, including
tree-launching, to hydra in the future release. -
Added in-repository wiki documentation.
-
Added stream workq to support optimizations for enqueue operations.
-
Better support for large count APIs by eliminating type conversion issues.
-
Hydra now uses libpmi (src/pmi) for handling PMI messages.
-
Many bug fixes and enhancements.