You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The MPI RMA port still does poorly with older releases of MPI. Currently,
OpenMPI 5.0.x and MPICH 4.2.x appear to do well.
Added
GA can be initialized from an MPI communicator. This enables instances of GA
to be created from non-GA programs and run independently on a subset of
processors
An option to use System V memory instead of POSIX shared memory was added to
the MPI progress ranks runtime. This option was added to get around a memory
limitation with POSIX shared memory only allows a maximum of about half the
available memory to used as shared memory. System V memory does not have
this limit, but this implementation does not appear to be reliable and
should be avoided if large memory usage is not needed
Substantial support for sparse 2D arrays. This includes operations such as
matrix-vector and matrix-matrix multiplies, sparse-dense conversion and
operations on the diagonals
Changed
Updated compiler settings in CMake build if Fujitsu compilers are detected.
Other general improvements have also been made in CMake build.
Removed all instances of ga_malloc from GA source code. This frees GA from
needing to initialize MA before using GA library calls.
Modified progress ranks runtime to use hostnames instead of host IDs to
identify ranks on same SMP node. This is a more robust implementation.
Fixed
Fixed gcc toolchain checks in CMake for clang build
Fixed tiled arrays so that they work with restricted arrays and fixed some
additional bugs in block cyclic distributions
Removed several memory leaks
Older TCGMSG implementations removed from build and documentation.
TCGMSG-MPI is still retained
A substantial amount of code and documentation for discontinued or
deprecated systems was removed. Examples include Blue Gene P/L/Q, LAPI,
PVM, CCA, VIA, Quadrics, Elan-3/4, Myrinet GM, Mellanox, Hitachi, DEC, etc.
Number of outstanding non-blocking messages in MPI RMA runtime modified so
that non-blocking test passes