@@ -4,9 +4,9 @@ Open MPI v5.0.x series
4
4
This file contains all the NEWS updates for the Open MPI v5.0.x
5
5
series, in reverse chronological order.
6
6
7
- Open MPI version 5.0.0rc10
7
+ Open MPI version 5.0.0rc11
8
8
--------------------------
9
- :Date: 2 February 2023
9
+ :Date: 6 April 2023
10
10
11
11
.. admonition :: MPIR API has been removed
12
12
:class: warning
@@ -40,32 +40,28 @@ Open MPI version 5.0.0rc10
40
40
libraries, rather than linked into the Open MPI core libraries.
41
41
42
42
43
- - Changes since rc9:
44
-
45
- - coll/ucc: Added support for Scatter and Iscatter collectives.
46
- - Added cache bypass mechanism to the ``OFI `` BTL. This fixes conflicts
47
- with Libfabric, which has its own registration cache. This addes a bypass
48
- flag which can be used for providers known to have their own registration cache.
49
- - common/ompio: implement pipelined read and write operation.
50
- This new new code path shows significant performance improvements for reading/writing
51
- device buffers compared to the previous implementation, and reduces the memory
52
- footprint of ``OMPIO `` by allocating smaller temporary buffers.
53
- - 32-bit builds have been disabled. Building Open MPI in a 32-bit environment
54
- is no longer supported.
55
- - MPI-4: MPI_Info_get() and MPI_Infi_get_valuelen() are now deprecated.
56
- - MPI-4: Issue a deprecation warning when MPI_Cancel() is called for a non-blocking send request.
57
- - Fixed various bugs encountered when running MPI under a debugger.
58
- - Cleaned up a number of memory leaks.
59
- - Cleaned up global symbol pollution that was leaking out of libmpi.
60
- - Removed opal_list_insert(), it was buggy and not used.
61
- Thanks to Jinyuan Wang for the contribution.
43
+ - Changes since rc10:
44
+
45
+ - The ``HAN `` collective is now enabled by default. This replaces ``tuned `` as the
46
+ default out-of-the-box collective component for Open MPI.
47
+ - Various fixes to make v5.0 ABI compatible with v4.1 compiled programs.
48
+ - Fixed support for ``OFI `` on RHEL7 and Libfabric < 1.9.
49
+ - Added the mca option ``--mca ompi_pml_base_check_pml 0|1 `` to skip
50
+ ``PML `` transport validation across processes. This can speed-up launch
51
+ times for users who know their cluster will always choose the same
52
+ ``PML `` transport. Default: verify pml selections.
53
+ - Use Libfabric 1.18 if available when using the ``OFI `` transport.
54
+ - Implemented ``ompi_info `` color coding.
55
+ - Added ``MPI_SESSION_NULL `` to Fortran bindings. Thanks to Jan Fecht for the fix.
56
+ - Fixed a bug where CUDA-aware MPI is broken when using the ``OB1 `` transport.
62
57
- Many other bug fixes and cleanups.
63
- - Various documentation updates.
58
+ - Many documentation updates.
59
+ Thanks to Nick Papior for the contributions.
64
60
65
61
- All other notable updates for v5.0.0:
66
62
67
- - Updated PMIx to the ``v4.2 `` branch - current hash: ``852d284 ``.
68
- - Updated PRRTE to the ``v3.0 `` branch - current hash: `` facdb35 ` `.
63
+ - Updated PMIx to the ``v4.2 `` branch - current hash: ``7d45393 ``.
64
+ - Updated PRRTE to the ``v3.0 `` branch - current hash: `20ee752 `.
69
65
70
66
- New Features:
71
67
@@ -93,6 +89,7 @@ Open MPI version 5.0.0rc10
93
89
- Added support for MPI minimum alignment key to the one-sided ``RDMA `` component.
94
90
- Add ability to detect patched memory to ``memory_patcher ``. Thanks
95
91
to Rich Welch for the contribution.
92
+ - coll/ucc: Added support for Scatter and Iscatter collectives.
96
93
97
94
- MPI-4.0 updates and additions:
98
95
@@ -111,6 +108,8 @@ Open MPI version 5.0.0rc10
111
108
- Droped unknown/ignored info keys on communicators, files,
112
109
and windows.
113
110
- Initial implementations of ``MPI_COMM_TYPE_HW_GUIDED `` and ``MPI_COMM_TYPE_HW_GUIDED `` added.
111
+ - ``MPI_Info_get() `` and ``MPI_Infi_get_valuelen() `` are now deprecated.
112
+ - Issue a deprecation warning when ``MPI_Cancel() `` is called for a non-blocking send request.
114
113
115
114
- Transport updates and improvements
116
115
@@ -131,7 +130,12 @@ Open MPI version 5.0.0rc10
131
130
interface matching between peers in order to improve ``MPI_Init() `` wireup
132
131
performance.
133
132
134
- - Changes to the BTL ``OFI `` component to better support the HPE SS11 network.
133
+ - OFI
134
+
135
+ - Improved support for the HPE SS11 network.
136
+ - Added cache bypass mechanism. This fixes conflicts
137
+ with Libfabric, which has its own registration cache. This addes a bypass
138
+ flag which can be used for providers known to have their own registration cache.
135
139
136
140
- Shared Memory:
137
141
@@ -164,6 +168,7 @@ Open MPI version 5.0.0rc10
164
168
This library was incomplete and unmaintained. If needed, it
165
169
is available in the v4/v4.1 series.
166
170
- The rankfile format no longer supports physical processor locations. Only logical processor locations are supported.
171
+ - 32-bit builds have been disabled. Building Open MPI in a 32-bit environment is no longer supported.
167
172
168
173
- HWLOC updates:
169
174
@@ -214,6 +219,10 @@ Open MPI version 5.0.0rc10
214
219
- Other updates and bug fixes:
215
220
216
221
- Updated Open MPI to use ``ROMIO `` v3.4.1.
222
+ - common/ompio: implement pipelined read and write operation.
223
+ This new new code path shows significant performance improvements for reading/writing
224
+ device buffers compared to the previous implementation, and reduces the memory
225
+ footprint of ``OMPIO `` by allocating smaller temporary buffers.
217
226
- Fixed Fortran-8-byte-INTEGER vs. C-4-byte-int issue in the ``mpi_f08 ``
218
227
MPI Fortran bindings module. Thanks to @ahaichen for reporting the bug.
219
228
- Add missing ``MPI_Status `` conversion subroutines:
0 commit comments