forked from open-mpi/ompi
-
Notifications
You must be signed in to change notification settings - Fork 1
/
README
2103 lines (1647 loc) · 86.5 KB
/
README
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
University Research and Technology
Corporation. All rights reserved.
Copyright (c) 2004-2015 The University of Tennessee and The University
of Tennessee Research Foundation. All rights
reserved.
Copyright (c) 2004-2008 High Performance Computing Center Stuttgart,
University of Stuttgart. All rights reserved.
Copyright (c) 2004-2007 The Regents of the University of California.
All rights reserved.
Copyright (c) 2006-2016 Cisco Systems, Inc. All rights reserved.
Copyright (c) 2006-2011 Mellanox Technologies. All rights reserved.
Copyright (c) 2006-2012 Oracle and/or its affiliates. All rights reserved.
Copyright (c) 2007 Myricom, Inc. All rights reserved.
Copyright (c) 2008 IBM Corporation. All rights reserved.
Copyright (c) 2010 Oak Ridge National Labs. All rights reserved.
Copyright (c) 2011 University of Houston. All rights reserved.
Copyright (c) 2013-2015 Intel, Inc. All rights reserved
Copyright (c) 2015 NVIDIA Corporation. All rights reserved.
$COPYRIGHT$
Additional copyrights may follow
$HEADER$
===========================================================================
When submitting questions and problems, be sure to include as much
extra information as possible. This web page details all the
information that we request in order to provide assistance:
http://www.open-mpi.org/community/help/
The best way to report bugs, send comments, or ask questions is to
sign up on the user's and/or developer's mailing list (for user-level
and developer-level questions; when in doubt, send to the user's
list):
Because of spam, only subscribers are allowed to post to these lists
(ensure that you subscribe with and post from exactly the same e-mail
address -- [email protected] is considered different than
[email protected]!). Visit these pages to subscribe to the
lists:
http://www.open-mpi.org/mailman/listinfo.cgi/users
http://www.open-mpi.org/mailman/listinfo.cgi/devel
Thanks for your time.
===========================================================================
Much, much more information is also available in the Open MPI FAQ:
http://www.open-mpi.org/faq/
===========================================================================
The following abbreviated list of release notes applies to this code
base as of this writing (April 2015):
General notes
-------------
- Open MPI now includes two public software layers: MPI and OpenSHMEM.
Throughout this document, references to Open MPI implicitly include
both of these layers. When distinction between these two layers is
necessary, we will reference them as the "MPI" and "OSHMEM" layers
respectively.
- OpenSHMEM is a collaborative effort between academia, industry, and
the U.S. Government to create a specification for a standardized API
for parallel programming in the Partitioned Global Address Space
(PGAS). For more information about the OpenSHMEM project, including
access to the current OpenSHMEM specification, please visit:
http://openshmem.org/
This OpenSHMEM implementation is provided on an experimental basis;
it has been lightly tested and will only work in Linux environments.
Although this implementation attempts to be portable to multiple
different environments and networks, it is still new and will likely
experience growing pains typical of any new software package.
End-user feedback is greatly appreciated.
This implementation will currently most likely provide optimal
performance on Mellanox hardware and software stacks. Overall
performance is expected to improve as other network vendors and/or
institutions contribute platform specific optimizations.
See below for details on how to enable the OpenSHMEM implementation.
- Open MPI includes support for a wide variety of supplemental
hardware and software package. When configuring Open MPI, you may
need to supply additional flags to the "configure" script in order
to tell Open MPI where the header files, libraries, and any other
required files are located. As such, running "configure" by itself
may not include support for all the devices (etc.) that you expect,
especially if their support headers / libraries are installed in
non-standard locations. Network interconnects are an easy example
to discuss -- Libfabric and OpenFabrics networks, for example, both
have supplemental headers and libraries that must be found before
Open MPI can build support for them. You must specify where these
files are with the appropriate options to configure. See the
listing of configure command-line switches, below, for more details.
- The majority of Open MPI's documentation is here in this file, the
included man pages, and on the web site FAQ
(http://www.open-mpi.org/).
- Note that Open MPI documentation uses the word "component"
frequently; the word "plugin" is probably more familiar to most
users. As such, end users can probably completely substitute the
word "plugin" wherever you see "component" in our documentation.
For what it's worth, we use the word "component" for historical
reasons, mainly because it is part of our acronyms and internal API
function calls.
- The run-time systems that are currently supported are:
- rsh / ssh
- LoadLeveler
- PBS Pro, Torque
- Platform LSF (v7.0.2 and later)
- SLURM
- Cray XE, XC, and XK
- Oracle Grid Engine (OGE) 6.1, 6.2 and open source Grid Engine
- Systems that have been tested are:
- Linux (various flavors/distros), 32 bit, with gcc
- Linux (various flavors/distros), 64 bit (x86), with gcc, Absoft,
Intel, and Portland (*)
- OS X (10.6, 10.7, 10.8, 10.9, 10.10), 32 and 64 bit (x86_64), with
XCode and Absoft compilers (*)
(*) Be sure to read the Compiler Notes, below.
- Other systems have been lightly (but not fully tested):
- Cygwin 32 & 64 bit with gcc
- ARMv4, ARMv5, ARMv6, ARMv7, ARMv8
- Other 64 bit platforms (e.g., Linux on PPC64)
- Oracle Solaris 10 and 11, 32 and 64 bit (SPARC, i386, x86_64),
with Oracle Solaris Studio 12.2, 12.3, and 12.4
Compiler Notes
--------------
- Open MPI requires a C99-capable compiler to build.
- Mixing compilers from different vendors when building Open MPI
(e.g., using the C/C++ compiler from one vendor and the Fortran
compiler from a different vendor) has been successfully employed by
some Open MPI users (discussed on the Open MPI user's mailing list),
but such configurations are not tested and not documented. For
example, such configurations may require additional compiler /
linker flags to make Open MPI build properly.
- In general, the latest versions of compilers of a given vendor's
series have the least bugs. We have seen cases where Vendor XYZ's
compiler version A.B fails to compile Open MPI, but version A.C
(where C>B) works just fine. If you run into a compile failure, you
might want to double check that you have the latest bug fixes and
patches for your compiler.
- Users have reported issues with older versions of the Fortran PGI
compiler suite when using Open MPI's (non-default) --enable-debug
configure option. Per the above advice of using the most recent
version of a compiler series, the Open MPI team recommends using the
latest version of the PGI suite, and/or not using the --enable-debug
configure option. If it helps, here's what we have found with some
(not comprehensive) testing of various versions of the PGI compiler
suite:
pgi-8 : NO known good version with --enable-debug
pgi-9 : 9.0-4 known GOOD
pgi-10: 10.0-0 known GOOD
pgi-11: NO known good version with --enable-debug
pgi-12: 12.10 known BAD due to C99 compliance issues, and 12.8
and 12.9 both known BAD with --enable-debug
pgi-13: 13.10 known GOOD
- Similarly, there is a known Fortran PGI compiler issue with long
source directory path names that was resolved in 9.0-4 (9.0-3 is
known to be broken in this regard).
- IBM's xlf compilers: NO known good version that can build/link
the MPI f08 bindings or build/link the OSHMEM Fortran bindings.
- On NetBSD-6 (at least AMD64 and i386), and possibly on OpenBSD,
libtool misidentifies properties of f95/g95, leading to obscure
compile-time failures if used to build Open MPI. You can work
around this issue by ensuring that libtool will not use f95/g95
(e.g., by specifying FC=<some_other_compiler>, or otherwise ensuring
a different Fortran compiler will be found earlier in the path than
f95/g95), or by disabling the Fortran MPI bindings with
--disable-mpi-fortran.
- Absoft 11.5.2 plus a service pack from September 2012 (which Absoft
says is available upon request), or a version later than 11.5.2
(e.g., 11.5.3), is required to compile the new Fortran mpi_f08
module.
- Open MPI does not support the Sparc v8 CPU target. However,
as of Solaris Studio 12.1, and later compilers, one should not
specify -xarch=v8plus or -xarch=v9. The use of the options
-m32 and -m64 for producing 32 and 64 bit targets, respectively,
are now preferred by the Solaris Studio compilers. GCC may
require either "-m32" or "-mcpu=v9 -m32", depending on GCC version.
- It has been noticed that if one uses CXX=sunCC, in which sunCC
is a link in the Solaris Studio compiler release, that the OMPI
build system has issue with sunCC and does not build libmpi_cxx.so.
Therefore the make install fails. So we suggest that one should
use CXX=CC, which works, instead of CXX=sunCC.
- If one tries to build OMPI on Ubuntu with Solaris Studio using the C++
compiler and the -m32 option, you might see a warning:
CC: Warning: failed to detect system linker version, falling back to
custom linker usage
And the build will fail. One can overcome this error by either
setting LD_LIBRARY_PATH to the location of the 32 bit libraries (most
likely /lib32), or giving LDFLAGS="-L/lib32 -R/lib32" to the configure
command. Officially, Solaris Studio is not supported on Ubuntu Linux
distributions, so additional problems might be incurred.
- Open MPI does not support the gccfss compiler (GCC For SPARC
Systems; a now-defunct compiler project from Sun).
- At least some versions of the Intel 8.1 compiler seg fault while
compiling certain Open MPI source code files. As such, it is not
supported.
- The Intel 9.0 v20051201 compiler on IA64 platforms seems to have a
problem with optimizing the ptmalloc2 memory manager component (the
generated code will segv). As such, the ptmalloc2 component will
automatically disable itself if it detects that it is on this
platform/compiler combination. The only effect that this should
have is that the MCA parameter mpi_leave_pinned will be inoperative.
- It has been reported that the Intel 9.1 and 10.0 compilers fail to
compile Open MPI on IA64 platforms. As of 12 Sep 2012, there is
very little (if any) testing performed on IA64 platforms (with any
compiler). Support is "best effort" for these platforms, but it is
doubtful that any effort will be expended to fix the Intel 9.1 /
10.0 compiler issuers on this platform.
- Early versions of the Intel 12.1 Linux compiler suite on x86_64 seem
to have a bug that prevents Open MPI from working. Symptoms
including immediate segv of the wrapper compilers (e.g., mpicc) and
MPI applications. As of 1 Feb 2012, if you upgrade to the latest
version of the Intel 12.1 Linux compiler suite, the problem will go
away.
- Early versions of the Portland Group 6.0 compiler have problems
creating the C++ MPI bindings as a shared library (e.g., v6.0-1).
Tests with later versions show that this has been fixed (e.g.,
v6.0-5).
- The Portland Group compilers prior to version 7.0 require the
"-Msignextend" compiler flag to extend the sign bit when converting
from a shorter to longer integer. This is is different than other
compilers (such as GNU). When compiling Open MPI with the Portland
compiler suite, the following flags should be passed to Open MPI's
configure script:
shell$ ./configure CFLAGS=-Msignextend CXXFLAGS=-Msignextend \
--with-wrapper-cflags=-Msignextend \
--with-wrapper-cxxflags=-Msignextend ...
This will both compile Open MPI with the proper compile flags and
also automatically add "-Msignextend" when the C and C++ MPI wrapper
compilers are used to compile user MPI applications.
- Using the MPI C++ bindings with older versions of the Pathscale
compiler on some platforms is an old issue that seems to be a
problem when Pathscale uses a back-end GCC 3.x compiler. Here's a
proposed solution from the Pathscale support team (from July 2010):
The proposed work-around is to install gcc-4.x on the system and
use the pathCC -gnu4 option. Newer versions of the compiler (4.x
and beyond) should have this fixed, but we'll have to test to
confirm it's actually fixed and working correctly.
We don't anticipate that this will be much of a problem for Open MPI
users these days (our informal testing shows that not many users are
still using GCC 3.x). Contact Pathscale support if you continue to
have problems with Open MPI's C++ bindings.
- Using the Absoft compiler to build the MPI Fortran bindings on Suse
9.3 is known to fail due to a Libtool compatibility issue.
- MPI Fortran API support has been completely overhauled since the
Open MPI v1.5/v1.6 series.
********************************************************************
********************************************************************
*** There is now only a single Fortran MPI wrapper compiler and a
*** single Fortran OSHMEM wrapper compiler: mpifort and oshfort,
*** respectively. mpif77 and mpif90 still exist, but they are
*** symbolic links to mpifort.
********************************************************************
*** Similarly, Open MPI's configure script only recognizes the FC
*** and FCFLAGS environment variables (to specify the Fortran
*** compiler and compiler flags, respectively). The F77 and FFLAGS
*** environment variables are IGNORED.
********************************************************************
********************************************************************
As a direct result, it is STRONGLY recommended that you specify a
Fortran compiler that uses file suffixes to determine Fortran code
layout (e.g., free form vs. fixed). For example, with some versions
of the IBM XLF compiler, it is preferable to use FC=xlf instead of
FC=xlf90, because xlf will automatically determine the difference
between free form and fixed Fortran source code.
However, many Fortran compilers allow specifying additional
command-line arguments to indicate which Fortran dialect to use.
For example, if FC=xlf90, you may need to use "mpifort --qfixed ..."
to compile fixed format Fortran source files.
You can use either ompi_info or oshmem_info to see with which Fortran
compiler Open MPI was configured and compiled.
There are up to three sets of Fortran MPI bindings that may be
provided depending on your Fortran compiler):
- mpif.h: This is the first MPI Fortran interface that was defined
in MPI-1. It is a file that is included in Fortran source code.
Open MPI's mpif.h does not declare any MPI subroutines; they are
all implicit.
- mpi module: The mpi module file was added in MPI-2. It provides
strong compile-time parameter type checking for MPI subroutines.
- mpi_f08 module: The mpi_f08 module was added in MPI-3. It
provides many advantages over the mpif.h file and mpi module. For
example, MPI handles have distinct types (vs. all being integers).
See the MPI-3 document for more details.
*** The mpi_f08 module is STRONGLY is recommended for all new MPI
Fortran subroutines and applications. Note that the mpi_f08
module can be used in conjunction with the other two Fortran
MPI bindings in the same application (only one binding can be
used per subroutine/function, however). Full interoperability
between mpif.h/mpi module and mpi_f08 module MPI handle types
is provided, allowing mpi_f08 to be used in new subroutines in
legacy MPI applications.
Per the OSHMEM specification, there is only one Fortran OSHMEM binding
provided:
- shmem.fh: All Fortran OpenSHMEM programs **should** include 'shmem.fh',
and Fortran OSHMEM programs that use constants defined by OpenSHMEM
**MUST** include 'shmem.fh'.
The following notes apply to the above-listed Fortran bindings:
- All Fortran compilers support the mpif.h/shmem.fh-based bindings,
with one exception: the MPI_SIZEOF interfaces will only be present
when Open MPI is built with a Fortran compiler that support the
INTERFACE keyword and ISO_FORTRAN_ENV. Most notably, this
excludes the GNU Fortran compiler suite before version 4.9.
- The level of support provided by the mpi module is based on your
Fortran compiler.
If Open MPI is built with a non-GNU Fortran compiler, or if Open
MPI is built with the GNU Fortran compiler >= v4.9, all MPI
subroutines will be prototyped in the mpi module. All calls to
MPI subroutines will therefore have their parameter types checked
at compile time.
If Open MPI is built with an old gfortran (i.e., < v4.9), a
limited "mpi" module will be built. Due to the limitations of
these compilers, and per guidance from the MPI-3 specification,
all MPI subroutines with "choice" buffers are specifically *not*
included in the "mpi" module, and their parameters will not be
checked at compile time. Specifically, all MPI subroutines with
no "choice" buffers are prototyped and will receive strong
parameter type checking at run-time (e.g., MPI_INIT,
MPI_COMM_RANK, etc.).
Similar to the mpif.h interface, MPI_SIZEOF is only supported on
Fortran compilers that support INTERFACE and ISO_FORTRAN_ENV.
- The mpi_f08 module is new and has been tested with the Intel
Fortran compiler and gfortran >= 4.9. Other modern Fortran
compilers may also work (but are, as yet, only lightly tested).
It is expected that this support will mature over time.
Many older Fortran compilers do not provide enough modern Fortran
features to support the mpi_f08 module. For example, gfortran <
v4.9 does provide enough support for the mpi_f08 module.
You can examine the output of the following command to see all
the Fortran features that are/are not enabled in your Open MPI
installation:
shell$ ompi_info | grep -i fort
General Run-Time Support Notes
------------------------------
- The Open MPI installation must be in your PATH on all nodes (and
potentially LD_LIBRARY_PATH (or DYLD_LIBRARY_PATH), if libmpi/libshmem
is a shared library), unless using the --prefix or
--enable-mpirun-prefix-by-default functionality (see below).
- Open MPI's run-time behavior can be customized via MCA ("MPI
Component Architecture") parameters (see below for more information
on how to get/set MCA parameter values). Some MCA parameters can be
set in a way that renders Open MPI inoperable (see notes about MCA
parameters later in this file). In particular, some parameters have
required options that must be included.
- If specified, the "btl" parameter must include the "self"
component, or Open MPI will not be able to deliver messages to the
same rank as the sender. For example: "mpirun --mca btl tcp,self
..."
- If specified, the "btl_tcp_if_exclude" paramater must include the
loopback device ("lo" on many Linux platforms), or Open MPI will
not be able to route MPI messages using the TCP BTL. For example:
"mpirun --mca btl_tcp_if_exclude lo,eth1 ..."
- Running on nodes with different endian and/or different datatype
sizes within a single parallel job is supported in this release.
However, Open MPI does not resize data when datatypes differ in size
(for example, sending a 4 byte MPI_DOUBLE and receiving an 8 byte
MPI_DOUBLE will fail).
MPI Functionality and Features
------------------------------
- Rank reordering support is available using the TreeMatch library. It
is activated for the graph and dist_graph topologies.
- All MPI-3 functionality is supported.
- When using MPI deprecated functions, some compilers will emit
warnings. For example:
shell$ cat deprecated_example.c
#include <mpi.h>
void foo(void) {
MPI_Datatype type;
MPI_Type_struct(1, NULL, NULL, NULL, &type);
}
shell$ mpicc -c deprecated_example.c
deprecated_example.c: In function 'foo':
deprecated_example.c:4: warning: 'MPI_Type_struct' is deprecated (declared at /opt/openmpi/include/mpi.h:1522)
shell$
- MPI_THREAD_MULTIPLE support is included, but is only lightly tested.
It likely does not work for thread-intensive applications. Note
that *only* the MPI point-to-point communication functions for the
BTL's listed here are considered thread safe. Other support
functions (e.g., MPI attributes) have not been certified as safe
when simultaneously used by multiple threads.
- tcp
- sm
- self
Note that Open MPI's thread support is in a fairly early stage; the
above devices may *work*, but the latency is likely to be fairly
high. Specifically, efforts so far have concentrated on
*correctness*, not *performance* (yet).
YMMV.
- MPI_REAL16 and MPI_COMPLEX32 are only supported on platforms where a
portable C datatype can be found that matches the Fortran type
REAL*16, both in size and bit representation.
- The "libompitrace" library is bundled in Open MPI and is installed
by default (it can be disabled via the --disable-libompitrace
flag). This library provides a simplistic tracing of select MPI
function calls via the MPI profiling interface. Linking it in to
your appliation via (e.g., via -lompitrace) will automatically
output to stderr when some MPI functions are invoked:
shell$ cd examples/
shell$ mpicc hello_c.c -o hello_c -lompitrace
shell$ mpirun -np 1 hello_c
MPI_INIT: argc 1
Hello, world, I am 0 of 1
MPI_BARRIER[0]: comm MPI_COMM_WORLD
MPI_FINALIZE[0]
shell$
Keep in mind that the output from the trace library is going to
stderr, so it may output in a slightly different order than the
stdout from your application.
This library is being offered as a "proof of concept" / convenience
from Open MPI. If there is interest, it is trivially easy to extend
it to printf for other MPI functions. Patches and/or suggestions
would be greatfully appreciated on the Open MPI developer's list.
OSHMEM Functionality and Features
------------------------------
- All OpenSHMEM-1.0 functionality is supported.
MPI Collectives
-----------
- The "hierarch" coll component (i.e., an implementation of MPI
collective operations) attempts to discover network layers of
latency in order to segregate individual "local" and "global"
operations as part of the overall collective operation. In this
way, network traffic can be reduced -- or possibly even minimized
(similar to MagPIe). The current "hierarch" component only
separates MPI processes into on- and off-node groups.
Hierarch has had sufficient correctness testing, but has not
received much performance tuning. As such, hierarch is not
activated by default -- it must be enabled manually by setting its
priority level to 100:
mpirun --mca coll_hierarch_priority 100 ...
We would appreciate feedback from the user community about how well
hierarch works for your applications.
- The "fca" coll component: the Mellanox Fabric Collective Accelerator
(FCA) is a solution for offloading collective operations from the
MPI process onto Mellanox QDR InfiniBand switch CPUs and HCAs.
- The "ML" coll component is an implementation of MPI collective
operations that takes advantage of communication hierarchies in
modern systems. A ML collective operation is implemented by
combining multiple independently progressing collective primitives
implemented over different communication hierarchies, hence a ML
collective operation is also referred to as a hierarchical
collective operation. The number of collective primitives that are
included in a ML collective operation is a function of
subgroups(hierarchies). Typically, MPI processes in a single
communication hierarchy such as CPU socket, node, or subnet are
grouped together into a single subgroup (hierarchy). The number of
subgroups are configurable at runtime, and each different collective
operation could be configured to have a different of number of
subgroups.
The component frameworks and components used by/required for a
"ML" collective operation.
Frameworks:
* "sbgp" - Provides functionality for grouping processes into
subgroups
* "bcol" - Provides collective primitives optimized for a particular
communication hierarchy
Components:
* sbgp components - Provides grouping functionality over a CPU
socket ("basesocket"), shared memory
("basesmuma"), Mellanox's ConnectX HCA
("ibnet"), and other interconnects supported by
PML ("p2p")
* BCOL components - Provides optimized collective primitives for
shared memory ("basesmuma"), Mellanox's ConnectX
HCA ("iboffload"), and other interconnects
supported by PML ("ptpcoll")
- The "cuda" coll component provides CUDA-aware support for the
reduction type collectives with GPU buffers. This component is only
compiled into the library when the library has been configured with
CUDA-aware support. It intercepts calls to the reduction
collectives, copies the data to staging buffers if GPU buffers, then
calls underlying collectives to do the work.
OSHMEM Collectives
-----------
- The "fca" scoll component: the Mellanox Fabric Collective Accelerator
(FCA) is a solution for offloading collective operations from the
MPI process onto Mellanox QDR InfiniBand switch CPUs and HCAs.
- The "basic" scoll component: Reference implementation of all OSHMEM
collective operations.
Network Support
---------------
- There are three main MPI network models available: "ob1", "cm", and
"yalla". "ob1" uses BTL ("Byte Transfer Layer") components for each
supported network. "cm" uses MTL ("Matching Tranport Layer")
components for each supported network. "yalla" uses the Mellanox
MXM transport.
- "ob1" supports a variety of networks that can be used in
combination with each other (per OS constraints; e.g., there are
reports that the GM and OpenFabrics kernel drivers do not operate
well together):
- OpenFabrics: InfiniBand, iWARP, and RoCE
- Loopback (send-to-self)
- Shared memory
- TCP
- Intel Phi SCIF
- SMCUDA
- Cisco usNIC
- uGNI (Cray Gemini, Aries)
- vader (XPMEM, Linux CMA, Linux KNEM, and general shared memory)
- "cm" supports a smaller number of networks (and they cannot be
used together), but may provide better overall MPI performance:
- QLogic InfiniPath / Intel True Scale PSM
- Intel Omni-Path PSM2
- Mellanox MXM
- Portals4
- OpenFabrics Interfaces ("libfabric" tag matching)
Open MPI will, by default, choose to use "cm" when one of the
above transports can be used. Otherwise, "ob1" will be used and
the corresponding BTLs will be selected. Users can force the use
of ob1 or cm if desired by setting the "pml" MCA parameter at
run-time:
shell$ mpirun --mca pml ob1 ...
or
shell$ mpirun --mca pml cm ...
- Similarly, there are two OSHMEM network models available: "yoda",
and "ikrit". "yoda" also uses the BTL components for many supported
network. "ikrit" interfaces directly with Mellanox MXM.
- "yoda" supports a variety of networks that can be used:
- OpenFabrics: InfiniBand, iWARP, and RoCE
- Loopback (send-to-self)
- Shared memory
- TCP
- "ikrit" only supports Mellanox MXM.
- MXM is the Mellanox Messaging Accelerator library utilizing a full
range of IB transports to provide the following messaging services
to the upper level MPI/OSHMEM libraries:
- Usage of all available IB transports
- Native RDMA support
- Progress thread
- Shared memory communication
- Hardware-assisted reliability
- The usnic BTL is support for Cisco's usNIC device ("userspace NIC")
on Cisco UCS servers with the Virtualized Interface Card (VIC).
Although the usNIC is accessed via the OpenFabrics Libfabric API
stack, this BTL is specific to the Cisco usNIC device.
- uGNI is a Cray library for communicating over the Gemini and Aries
interconnects.
- The OpenFabrics Enterprise Distribution (OFED) software package v1.0
will not work properly with Open MPI v1.2 (and later) due to how its
Mellanox InfiniBand plugin driver is created. The problem is fixed
OFED v1.1 (and later).
- Better memory management support is available for OFED-based
transports using the "ummunotify" Linux kernel module. OFED memory
managers are necessary for better bandwidth when re-using the same
buffers for large messages (e.g., benchmarks and some applications).
Unfortunately, the ummunotify module was not accepted by the Linux
kernel community (and is still not distributed by OFED). But it
still remains the best memory management solution for MPI
applications that used the OFED network transports. If Open MPI is
able to find the <linux/ummunotify.h> header file, it will build
support for ummunotify and include it by default. If MPI processes
then find the ummunotify kernel module loaded and active, then their
memory managers (which have been shown to be problematic in some
cases) will be disabled and ummunotify will be used. Otherwise, the
same memory managers from prior versions of Open MPI will be used.
The ummunotify Linux kernel module can be downloaded from:
http://lwn.net/Articles/343351/
- The use of fork() with OpenFabrics-based networks (i.e., the openib
BTL) is only partially supported, and only on Linux kernels >=
v2.6.15 with libibverbs v1.1 or later (first released as part of
OFED v1.2), per restrictions imposed by the OFED network stack.
- Linux "knem" support is used when the "vader" or "sm" (shared
memory) BTLs are compiled with knem support (see the --with-knem
configure option) and the knem Linux module is loaded in the running
kernel. If the knem Linux kernel module is not loaded, the knem
support is (by default) silently deactivated during Open MPI jobs.
See http://runtime.bordeaux.inria.fr/knem/ for details on Knem.
- Linux Cross-Memory Attach (CMA) or XPMEM is used by the vader
shared-memory BTL when the CMA/XPMEM libraries are installedm,
respectively. Linux CMA and XPMEM are similar (but different)
mechanisms for Open MPI to utilize single-copy semantics for shared
memory.
Open MPI Extensions
-------------------
- An MPI "extensions" framework has been added (but is not enabled by
default). See the "Open MPI API Extensions" section below for more
information on compiling and using MPI extensions.
- The following extensions are included in this version of Open MPI:
- affinity: Provides the OMPI_Affinity_str() routine on retrieving
a string that contains what resources a process is bound to. See
its man page for more details.
- cr: Provides routines to access to checkpoint restart routines.
See ompi/mpiext/cr/mpiext_cr_c.h for a listing of availble
functions.
- cuda: When the library is compiled with CUDA-aware support, it provides
two things. First, a macro MPIX_CUDA_AWARE_SUPPORT. Secondly, the
function MPIX_Query_cuda_support that can be used to query for support.
- example: A non-functional extension; its only purpose is to
provide an example for how to create other extensions.
===========================================================================
Building Open MPI
-----------------
Open MPI uses a traditional configure script paired with "make" to
build. Typical installs can be of the pattern:
---------------------------------------------------------------------------
shell$ ./configure [...options...]
shell$ make all install
---------------------------------------------------------------------------
There are many available configure options (see "./configure --help"
for a full list); a summary of the more commonly used ones is included
below.
Note that for many of Open MPI's --with-<foo> options, Open MPI will,
by default, search for header files and/or libraries for <foo>. If
the relevant files are found, Open MPI will built support for <foo>;
if they are not found, Open MPI will skip building support for <foo>.
However, if you specify --with-<foo> on the configure command line and
Open MPI is unable to find relevant support for <foo>, configure will
assume that it was unable to provide a feature that was specifically
requested and will abort so that a human can resolve out the issue.
INSTALLATION OPTIONS
--prefix=<directory>
Install Open MPI into the base directory named <directory>. Hence,
Open MPI will place its executables in <directory>/bin, its header
files in <directory>/include, its libraries in <directory>/lib, etc.
--disable-shared
By default, libmpi and libshmem are built as a shared library, and
all components are built as dynamic shared objects (DSOs). This
switch disables this default; it is really only useful when used with
--enable-static. Specifically, this option does *not* imply
--enable-static; enabling static libraries and disabling shared
libraries are two independent options.
--enable-static
Build libmpi and libshmem as static libraries, and statically link in all
components. Note that this option does *not* imply
--disable-shared; enabling static libraries and disabling shared
libraries are two independent options.
Be sure to read the description of --without-memory-manager, below;
it may have some effect on --enable-static.
--disable-wrapper-rpath
By default, the wrapper compilers (e.g., mpicc) will enable "rpath"
support in generated executables on systems that support it. That
is, they will include a file reference to the location of Open MPI's
libraries in the application executable itself. This means that
the user does not have to set LD_LIBRARY_PATH to find Open MPI's
libraries (e.g., if they are installed in a location that the
run-time linker does not search by default).
On systems that utilize the GNU ld linker, recent enough versions
will actually utilize "runpath" functionality, not "rpath". There
is an important difference between the two:
"rpath": the location of the Open MPI libraries is hard-coded into
the MPI/OSHMEM application and cannot be overridden at run-time.
"runpath": the location of the Open MPI libraries is hard-coded into
the MPI/OSHMEM application, but can be overridden at run-time by
setting the LD_LIBRARY_PATH environment variable.
For example, consider that you install Open MPI vA.B.0 and
compile/link your MPI/OSHMEM application against it. Later, you install
Open MPI vA.B.1 to a different installation prefix (e.g.,
/opt/openmpi/A.B.1 vs. /opt/openmpi/A.B.0), and you leave the old
installation intact.
In the rpath case, your MPI application will always use the
libraries from your A.B.0 installation. In the runpath case, you
can set the LD_LIBRARY_PATH environment variable to point to the
A.B.1 installation, and then your MPI application will use those
libraries.
Note that in both cases, however, if you remove the original A.B.0
installation and set LD_LIBRARY_PATH to point to the A.B.1
installation, your application will use the A.B.1 libraries.
This rpath/runpath behavior can be disabled via
--disable-wrapper-rpath.
--enable-dlopen
Build all of Open MPI's components as standalone Dynamic Shared
Objects (DSO's) that are loaded at run-time (this is the default).
The opposite of this option, --disable-dlopen, causes two things:
1. All of Open MPI's components will be built as part of Open MPI's
normal libraries (e.g., libmpi).
2. Open MPI will not attempt to open any DSO's at run-time.
Note that this option does *not* imply that OMPI's libraries will be
built as static objects (e.g., libmpi.a). It only specifies the
location of OMPI's components: standalone DSOs or folded into the
Open MPI libraries. You can control whether Open MPI's libraries
are build as static or dynamic via --enable|disable-static and
--enable|disable-shared.
--with-platform=FILE
Load configure options for the build from FILE. Options on the
command line that are not in FILE are also used. Options on the
command line and in FILE are replaced by what is in FILE.
NETWORKING SUPPORT / OPTIONS
--with-fca=<directory>
Specify the directory where the Mellanox FCA library and
header files are located.
FCA is the support library for Mellanox QDR switches and HCAs.
--with-hcoll=<directory>
Specify the directory where the Mellanox hcoll library and header
files are located. This option is generally only necessary if the
hcoll headers and libraries are not in default compiler/linker
search paths.
hcoll is the support library for MPI collective operation offload on
Mellanox ConnectX-3 HCAs (and later).
--with-knem=<directory>
Specify the directory where the knem libraries and header files are
located. This option is generally only necessary if the knem headers
and libraries are not in default compiler/linker search paths.
knem is a Linux kernel module that allows direct process-to-process
memory copies (optionally using hardware offload), potentially
increasing bandwidth for large messages sent between messages on the
same server. See http://runtime.bordeaux.inria.fr/knem/ for
details.
--with-libfabric=<directory>
Specify the directory where the OpenFabrics Interfaces libfabric
library and header files are located. This option is generally only
necessary if the libfabric headers and libraries are not in default
compiler/linker search paths.
Libfabric is the support library for OpenFabrics Interfaces-based
network adapters, such as Cisco usNIC, Intel True Scale PSM, etc.
--with-libfabric-libdir=<directory>
Look in directory for the libfabric libraries. By default, Open MPI
will look in <libfabric directory>/lib and <libfabric
directory>/lib64, which covers most cases. This option is only
needed for special configurations.
--with-mxm=<directory>
Specify the directory where the Mellanox MXM library and header
files are located. This option is generally only necessary if the
MXM headers and libraries are not in default compiler/linker search
paths.
MXM is the support library for Mellanox Network adapters.
--with-mxm-libdir=<directory>
Look in directory for the MXM libraries. By default, Open MPI will
look in <mxm directory>/lib and <mxm directory>/lib64, which covers
most cases. This option is only needed for special configurations.
--with-portals4=<directory>
Specify the directory where the Portals4 libraries and header files
are located. This option is generally only necessary if the Portals4
headers and libraries are not in default compiler/linker search
paths.
Portals is a low-level network API for high-performance networking
on high-performance computing systems developed by Sandia National
Laboratories, Intel Corporation, and the University of New Mexico.
The Portals 4 Reference Implementation is a complete implementation
of Portals 4, with transport over InfiniBand verbs and UDP.
--with-portals4-libdir=<directory>
Location of libraries to link with for Portals4 support.
--with-portals4-max-md-size=SIZE
--with-portals4-max-va-size=SIZE
Set configuration values for Portals 4
--with-psm=<directory>
Specify the directory where the QLogic InfiniPath / Intel True Scale
PSM library and header files are located. This option is generally
only necessary if the PSM headers and libraries are not in default
compiler/linker search paths.
PSM is the support library for QLogic InfiniPath and Intel TrueScale
network adapters.
--with-psm-libdir=<directory>
Look in directory for the PSM libraries. By default, Open MPI will
look in <psm directory>/lib and <psm directory>/lib64, which covers
most cases. This option is only needed for special configurations.
--with-psm2=<directory>
Specify the directory where the Intel Omni-Path PSM2 library and
header files are located. This option is generally only necessary
if the PSM2 headers and libraries are not in default compiler/linker
search paths.
PSM2 is the support library for Intel Omni-Path network adapters.
--with-psm2-libdir=<directory>
Look in directory for the PSM2 libraries. By default, Open MPI will
look in <psm2 directory>/lib and <psm2 directory>/lib64, which
covers most cases. This option is only needed for special
configurations.
--with-scif=<dir>
Look in directory for Intel SCIF support libraries
--with-verbs=<directory>
Specify the directory where the verbs (also know as OpenFabrics, and
previously known as OpenIB) libraries and header files are located.
This option is generally only necessary if the verbs headers and
libraries are not in default compiler/linker search paths.
"OpenFabrics" refers to operating system bypass networks, such as
InfiniBand, usNIC, iWARP, and RoCE (aka "IBoIP").
--with-verbs-libdir=<directory>
Look in directory for the verbs libraries. By default, Open MPI
will look in <verbs_directory>/lib and <verbs_ directory>/lib64,
which covers most cases. This option is only needed for special
configurations.
--with-verbs-usnic
This option will activate support in Open MPI for disabling a
dire-sounding warning message from libibverbs that Cisco usNIC
devices are not supported (because Cisco usNIC devices are supported
through libfabric, not libibverbs). This libibverbs warning can
also be suppressed by installing the "no op" libusnic_verbs plugin
for libibverbs (see https://github.com/cisco/libusnic_verbs, or
download binaries from cisco.com). This option is disabled by
default because it causes libopen-pal.so to depend on libibverbs.so,
which is undesirable to many downstream packagers.
--with-usnic
Abort configure if Cisco usNIC support cannot be built.
RUN-TIME SYSTEM SUPPORT
--enable-mpirun-prefix-by-default
This option forces the "mpirun" command to always behave as if
"--prefix $prefix" was present on the command line (where $prefix is
the value given to the --prefix option to configure). This prevents
most rsh/ssh-based users from needing to modify their shell startup
files to set the PATH and/or LD_LIBRARY_PATH for Open MPI on remote
nodes. Note, however, that such users may still desire to set PATH
-- perhaps even in their shell startup files -- so that executables
such as mpicc and mpirun can be found without needing to type long
path names. --enable-orterun-prefix-by-default is a synonym for
this option.
--enable-sensors
Enable internal sensors (default: disabled).
--enable-orte-static-ports
Enable orte static ports for tcp oob (default: enabled).
--with-alps
Force the building of for the Cray Alps run-time environment. If
Alps support cannot be found, configure will abort.
--with-loadleveler
Force the building of LoadLeveler scheduler support. If LoadLeveler
support cannot be found, configure will abort.
--with-lsf=<directory>
Specify the directory where the LSF libraries and header files are
located. This option is generally only necessary if the LSF headers