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
|
.. _label-building-fully-static-apps:
Building fully-static MPI applications
======================================
Is fully-static really what you need?
-------------------------------------
No one ends up in this documentation section by accident.
If you are reading this text, it is likely because you are looking to
solve a problem, and fully-static MPI applications sound like the
right answer. There are two common problems that people think that
fully-static MPI applications will solve:
#. MPI applications fail to launch because dependent shared libraries
are not found at run-time.
#. Filesystem performance when launching at scale is terrible.
If either of these are your problems, the Open MPI community *strongly
encourages you to use other mechanisms to fix the problem*:
fully-static MPI applications are *possible*, but are sub-optimal for
other reasons.
The following sections discuss the above problems.
Finding dependent shared libraries at run-time
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
If you are running in to problems finding shared libraries at run time
|mdash| particularly on nodes that are remote from where you have invoked
``mpirun`` |mdash| your best bet is to set ``LD_LIBRARY_PATH`` (or
equivalent mechanism) properly on *all* nodes (including remote
nodes).
This is an involved topic, but even so, it is generally simpler to
solve this problem than creating and maintaining static builds. See
:ref:`the section on Running MPI applications
<label-running-mpi-applications>` for more details.
Improving filesystem performance at scale
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Prior to v5.0.0, Open MPI compiled a large number of plugins as
individual dynamic shared objects (DSOs) |mdash| i.e., individual
files in the filesystem. Many of these DSOs would be opened by each
MPI process at run time.
This could cause filesystem congestion, particularly when Open MPI is
installed on a network filesystem and a large job is launched: many
nodes will simultaneously communicate with the file server(s), and
potentially need to transfer a large number of small(ish) files.
Starting with v5.0.0, by default, Open MPI's plugins are no longer
built as DSOs. As such, Open MPI typically only opens a small number
of shared libraries at launch time. Even if Open MPI is installed on
a network filesystem, these libraries are likely to be cached on nodes
over time, and therefore generate a fairly small amount network
filesystem traffic when MPI jobs are launched.
In short: Open MPI |ompi_ver|'s impact on network filesystems is
greatly diminished compared to prior versions. Compiling fully-static
applications to eliminate the open-every-DSO-file-at-launch-time
behavior is no longer necessary.
Other reasons fully-static applications are bad
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Here are a few other reasons that fully-static MPI applications are
sub-optimal:
#. When applications link all of their dependencies statically, the
operating system cannot share code between multiple copies of the
process.
For example, if you launch N copies of your fully-statically-linked
MPI application on a node, it will consume (N *
size_of_the_application) bytes of RAM. Alternately, launching N
copies of a dynamically-linked MPI application |mdash| where each
of the copies have the same dependent libraries |mdash| will only
load each shared dependent library into RAM once.
In other words: *using dynamic linking saves memory.*
#. *Fully*-static applications are not linked to the dynamic linking library,
e.g. ``libdl`` on Linux, which provides ``dlopen(3)``, ``dlsym(3)``, etc.
This will break Open MPI functionalities that depend on such interfaces.
.. warning:: Open MPI's memory management functionality, which provides
important performance optimizations on OS-bypass networks
such as InfiniBand, requires the ``dlsym(3)`` interface,
and therefore does not work with fully-static applications.
Are you convinced yet? *Please try to avoid building fully-static MPI
applications if at all possible.*
Building fully-static MPI applications
--------------------------------------
.. caution:: If, after reading all of the above, you are still of the
mind that you want to build fully-static MPI
applications, be aware that fully static linking is not
for the meek, and it is not recommended. But it is
possible, with some caveats.
#. You must have static libraries available for *everything* to which
your program links. This includes Open MPI; you must have used the
``--enable-static`` option to Open MPI's ``configure`` or otherwise
have available the static versions of the Open MPI libraries.
.. note:: Some Linux distributions may not have static versions of
popular Linux libraries by default (e.g., libnuma), or
require additional RPMs to be installed to get the
equivalent static libraries.
#. Open MPI must have been built without a memory manager. This means
that Open MPI must have been configured with the
``--without-memory-manager`` flag. This is irrelevant on some
platforms for which Open MPI does not have a memory manager, but on
some platforms it is necessary (Linux when using many OS-bypass
networks). It is harmless to use this flag on platforms where Open
MPI does not have a memory manager.
.. important:: Not including memory manager support can lead to
lower performance when Open MPI is used with
OS-bypass networks.
This is how to configure Open MPI to build fully-static libraries on
Linux:
.. code-block:: sh
shell$ ./configure --without-memory-manager --disable-dlopen \
--enable-static --disable-shared ...
The ``--disable-shared`` flag is optional; it will prevent Open MPI
from *also* building shared libraries.
Alternatively, you could build Open MPI with as many static libraries
as possible, but still preserve ``dlopen`` functionality by omitting
the ``--disable-dlopen`` flag:
.. code-block:: sh
shell$ ./configure --without-memory-manager \
--enable-static --disable-shared ...
This gives you a *mostly* static build of Open MPI, but has the
advantage of preserving at least some dynamic libraries.
Including whole archives
^^^^^^^^^^^^^^^^^^^^^^^^
Some systems may have additional constraints about their support
libraries that require additional steps to produce working
fully-static MPI applications. For example, any library that has its
own run-time plugin system (i.e., that opens dynamically shared
objects ("DSOs") at run time) will have additional complications in
producing fully-static builds.
In such cases, you generally want to run ``mpicc ... --showme`` to see
the compiler / linker commands that Open MPI's wrapper commands will
use, and then augment those commands with linker arguments for the
static versions of the DSO plugins that you will need at run time.
For example, if you have ``libfoo.a`` that dynamically loads
``plugin.so`` at run time, you'll need to have a ``plugin.a`` and
|mdash| assuming the GNU linker |mdash| add arguments similar to the
following:
* ``-static``: Tell the linker to generate a static executable.
* ``-Wl,--whole-archive -lfoo /path/to/plugin.a -Wl,--no-whole-archive``:
Tell the linker to include the entire ``foo`` library and the entire
``plugin.a`` archive in the executable.
You can either add these arguments on the command line manually, or
you can :ref:`modify the default behavior of the wrapper compilers
<label-customizing-wrapper-compiler>` to hide this complexity from end
users (but be aware that if you modify the wrapper compilers' default
behavior, *all* users will be creating static applications!).
|
