darshan-runtime.txt 17.2 KB
Newer Older
1 2 3 4 5
Darshan-runtime installation and usage
======================================

== Introduction

6 7 8
This document describes darshan-runtime, which is the instrumentation
portion of the Darshan characterization tool.  It should be installed on the
system where you intend to collect I/O characterization information.
Philip Carns's avatar
Philip Carns committed
9

10 11 12
Darshan instruments applications via either compile time wrappers for static
executables or dynamic library preloading for dynamic executables.  An
application that has been instrumented with Darshan will produce a single
Kevin Harms's avatar
Kevin Harms committed
13
log file each time it is executed.  This log summarizes the I/O access patterns
14 15 16 17 18 19 20 21 22 23
used by the application.

The darshan-runtime instrumentation only instruments MPI applications (the
application must at least call `MPI_Init()` and `MPI_Finalize()`).  However,
it captures both MPI-IO and POSIX file access.  It also captures limited
information about HDF5 and PnetCDF access.

This document provides generic installation instructions, but "recipes" for
several common HPC systems are provided at the end of the document as well.

24 25 26
More information about Darshan can be found at the 
http://www.mcs.anl.gov/darshan[Darshan web site].

27 28 29 30 31
== Requirements

* MPI C compiler
* zlib development headers and library

Philip Carns's avatar
Philip Carns committed
32
== Compilation
33 34 35 36 37 38 39 40 41 42

.Configure and build example
----
tar -xvzf darshan-<version-number>.tar.gz
cd darshan-<version-number>/darshan-runtime
./configure --with-mem-align=8 --with-log-path=/darshan-logs --with-jobid-env=PBS_JOBID CC=mpicc
make
make install
----

43 44 45 46 47 48 49 50 51 52 53 54 55 56
.Detecting file size and alignment
[NOTE]
====
You can also add --enable-stat-at-open option to cause the Darshan library
to issue an additional stat() system call on each file the first time that
it is opened on each process.  This allows Darshan to detect the file
alignment (and subsequent unaligned accesses).  It also allows Darshan to
detect the size of files at open time before any I/O is performed.
Unfortunately, this option can cause significant overhead at scale on file
systems such as PVFS or Lustre that must contact every server for a given
file in order to satisfy a stat request.  We therefore disable this
feature by default.
====

57 58 59 60 61 62 63 64 65 66 67 68 69 70
.Explanation of configure arguments:
* `--with-mem-align` (mandatory): This value is system-dependent and will be
used by Darshan to determine if the buffer for a read or write operation is
aligned in memory.
* `--with-log-path` (this, or `--with-log-path-by-env`, is mandatory): This
specifies the parent directory for the directory tree where darshan logs
will be placed
* `--with-jobid-env` (mandatory): this specifies the environment variable that
Darshan should check to determine the jobid of a job.  Common values are
`PBS_JOBID` or `COBALT_JOBID`.  If you are not using a scheduler (or your
scheduler does not advertise the job ID) then you can specify `NONE` here.
Darshan will fall back to using the pid of the rank 0 process if the
specified environment variable is not set.
* `CC=`: specifies the MPI C compiler to use for compilation
Philip Carns's avatar
Philip Carns committed
71
* `--with-log-path-by-env`: specifies an environment variable to use to
72 73 74 75
determine the log path at run time.
* `--with-log-hints=`: specifies hints to use when writing the Darshan log
file.  See `./configure --help` for details.
* `--with-zlib=`: specifies an alternate location for the zlib development
Philip Carns's avatar
Philip Carns committed
76
header and library.
77 78 79

=== Cross compilation

Philip Carns's avatar
Philip Carns committed
80
On some systems (notably the IBM Blue Gene series), the login nodes do not
81 82
have the same architecture or runtime environment as the compute nodes.  In
this case, you must configure darshan-runtime to be built using a cross
Philip Carns's avatar
Philip Carns committed
83
compiler.  The following configure arguments show an example for the BG/P system:
84 85 86 87 88 89 90

----
--host=powerpc-bgp-linux CC=/bgsys/drivers/ppcfloor/comm/default/bin/mpicc 
----

== Environment preparation

Philip Carns's avatar
Philip Carns committed
91 92
Once darshan-runtime has been installed, you must prepare a location
in which to store the Darshan log files and configure an instrumentation method.
93 94 95

=== Log directory

96
This step can be safely skipped if you configured darshan-runtime using the
Philip Carns's avatar
Philip Carns committed
97 98
`--with-log-path-by-env` option.  A more typical configuration uses a static
directory hierarchy for Darshan log
99 100 101 102 103
files.

The `darshan-mk-log-dirs.pl` utility will configure the path specified at
configure time to include
subdirectories organized by year, month, and day in which log files will be
Philip Carns's avatar
Philip Carns committed
104
placed. The deepest subdirectories will have sticky permissions to enable
105 106 107 108 109 110 111
multiple users to write to the same directory.  If the log directory is
shared system-wide across many users then the following script should be run
as root.
 
----
darshan-mk-log-dirs.pl
----
112

113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
.A note about log directory permissions
[NOTE]
====
All log files written by darshan have permissions set to only allow
read access by the owner of the file.  You can modify this behavior,
however, by specifying the --enable-group-readable-logs option at
configure time.  One notable deployment scenario would be to configure
Darshan and the log directories to allow all logs to be readable by both the
end user and a Darshan administrators group.   This can be done with the
following steps:

* set the --enable-group-readable-logs option at configure time
* create the log directories with darshan-mk-log-dirs.pl
* recursively set the group ownership of the log directories to the Darshan
administrators group
* recursively set the setgid bit on the log directories
====


132
=== Instrumentation method
133

134 135 136 137 138 139 140 141 142
The instrumentation method to use depends on whether the executables
produced by your MPI compiler are statically or dynamically linked.  If you
are unsure, you can check by running `ldd <executable_name>` on an example
executable.  Dynamically-linked executables will produce a list of shared
libraries when this command is executed.

Most MPI compilers allow you to toggle dynamic or static linking via options
such as `-dynamic` or `-static`.  Please check your MPI compiler man page
for details if you intend to force one mode or the other.
143 144 145

== Instrumenting statically-linked applications

146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162
Statically linked executables must be instrumented at compile time.  The
simplest way to do this is to generate an MPI compiler script (e.g. `mpicc`)
that includes the link options and libraries needed by Darshan.  Once this
is done, Darshan instrumentation is transparent; you simply compile
applications using the darshan-enabled MPI compiler scripts.

For MPICH-based MPI libraries, such as MPICH1, MPICH2, or MVAPICH, these
wrapper scripts can be generated automatically.  The following example
illustrates how to produce wrappers for C, C++, and Fortran compilers:

----
darshan-gen-cc.pl `which mpicc` --output mpicc.darshan
darshan-gen-cxx.pl `which mpicxx` --output mpicxx.darshan
darshan-gen-fortran.pl `which mpif77` --output mpif77.darshan
darshan-gen-fortran.pl `which mpif90` --output mpif90.darshan
-----

163 164 165
Please see the Cray recipe in this document for instructions on
instrumenting statically-linked applications on that platform.

166 167 168 169
For other MPI Libraries you must manually modify the MPI compiler scripts to
add the necessary link options and libraries.  Please see the
`darshan-gen-*` scripts for examples or contact the Darshan users mailing
list for help.
170 171 172

== Instrumenting dynamically-linked applications

173
For dynamically-linked executables, darshan relies on the `LD_PRELOAD`
Philip Carns's avatar
Philip Carns committed
174 175
environment variable to insert instrumentation at run time.  The executables
should be compiled using the normal, unmodified MPI compiler.
176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193

To use this mechanism, set the `LD_PRELOAD` environment variable to the full
path to the Darshan shared library, as in this example:

----
export LD_PRELOAD=/home/carns/darshan-install/lib/libdarshan.so
----

You can then run your application as usual.  Some environments may require a
special `mpirun` or `mpiexec` command line argument to propagate the
environment variable to all processes.  Other environments may require a
scheduler submission option to control this behavior.  Please check your
local site documentation for details.

=== Instrumenting dynamically-linked Fortran applications

Please follow the general steps outlined in the previous section.  For
Fortran applications compiled with MPICH you may have to take the additional
Philip Carns's avatar
Philip Carns committed
194
step of adding
195 196 197
`libfmpich.so` to your `LD_PRELOAD` environment variable. For example:

----
198
export LD_PRELOAD=/path/to/mpi/used/by/executable/lib/libfmpich.so:/home/carns/darshan-install/lib/libdarshan.so
199
----
200

201 202 203 204 205 206 207 208 209 210
[NOTE]
The full path to the libfmpich.so library can be omitted if the rpath
variable points to the correct path.  Be careful to check the rpath of the
darshan library and the executable before using this configuration, however.
They may provide conflicting paths.  Ideally the rpath to the  MPI library
would *not* be set by the darshan library, but would instead be specified
exclusively by the executable itself.  You can check the rpath of the
darshan library by running `objdump -x
/home/carns/darshan-install/lib/libdarshan.so |grep RPATH`.

211 212
== Darshan installation recipes

Philip Carns's avatar
Philip Carns committed
213
The following recipes provide examples for prominent HPC systems.
Philip Carns's avatar
Philip Carns committed
214
These are intended to be used as a starting point.  You will most likely have to adjust paths and options to
215 216 217 218 219 220 221 222 223 224 225 226 227 228
reflect the specifics of your system.

=== IBM Blue Gene/P

The IBM Blue Gene/P series produces static executables by default, uses a
different architecture for login and compute nodes, and uses an MPI
environment based on MPICH.

The following example shows how to configure Darshan on a BG/P system:

----
./configure --with-mem-align=16 \
 --with-log-path=/home/carns/working/darshan/releases/logs \
 --prefix=/home/carns/working/darshan/install --with-jobid-env=COBALT_JOBID \
229
 --with-zlib=/soft/apps/zlib-1.2.3/ \
230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
 --host=powerpc-bgp-linux CC=/bgsys/drivers/ppcfloor/comm/default/bin/mpicc 
----

.Rationale
[NOTE]
====
The memory alignment is set to 16 not because that is the proper alignment
for the BG/P CPU architecture, but because that is the optimal alignment for
the network transport used between compute nodes and I/O nodes in the
system.  The jobid environment variable is set to `COBALT_JOBID` in this
case for use with the Cobalt scheduler, but other BG/P systems may use
different schedulers.  The `--with-zlib` argument is used to point to a
version of zlib that has been compiled for use on the compute nodes rather
than the login node.  The `--host` argument is used to force cross-compilation
of Darshan.  The `CC` variable is set to point to a stock MPI compiler.
====

Once Darshan has been installed, use the `darshan-gen-*.pl` scripts as
described earlier in this document to produce darshan-enabled MPI compilers.
This method has been widely used and tested with both the GNU and IBM XL
compilers.

252
=== Cray platforms (XE, XC, or similar)
253

254 255 256 257 258 259 260
The Cray programming environment produces static executables by default,
which means that Darshan instrumentation must be inserted at compile
time.  This can be accomplished by loading a software module that sets
appropriate environment variables to modify the Cray compiler script link
behavior.  This section describes how to compile and install Darshan,
as well as how to use a software module to enable and disable Darshan
instrumentation.
261 262 263 264 265 266

==== Building and installing Darshan

Please set your environment to use the GNU programming environment before
configuring or compiling Darshan.  Although Darshan can be built with a
variety of compilers, the GNU compilers are recommended because it will
267 268 269 270 271
produce a Darshan library that is interoperable with the widest range
of compmilers and linkers.  On most Cray systems you can enable the GNU
programming environment with a command similar to "module swap PrgEnv-pgi
PrgEnv-gnu".  Please see your site documentation for information about
how to switch programming environments.
272 273

The following example shows how to configure and build Darshan on a Cray
274
system using either the GNU programming environment.  Adjust the 
275 276
--with-log-path and --prefix arguments to point to the desired log file path 
and installation path, respectively.
277 278

----
279
module swap PrgEnv-pgi PrgEnv-gnu
280
./configure --with-mem-align=8 \
281 282
 --with-log-path=/shared-file-system/darshan-logs \
 --prefix=/soft/darshan-2.2.3 \
283
 --with-jobid-env=PBS_JOBID --disable-cuserid CC=cc
284 285
make install
module swap PrgEnv-gnu PrgEnv-pgi
286 287 288 289 290 291 292
----

.Rationale
[NOTE]
====
The job ID is set to `PBS_JOBID` for use with a Torque or PBS based scheduler.
The `CC` variable is configured to point the standard MPI compiler.
Philip Carns's avatar
Philip Carns committed
293 294 295 296 297 298 299 300

The --disable-cuserid argument is used to prevent Darshan from attempting to
use the cuserid() function to retrieve the user name associated with a job.
Darshan automatically falls back to other methods if this function fails,
but on some Cray environments (notably the Beagle XE6 system as of March 2012)
the cuserid() call triggers a segmentation fault.  With this option set,
Darshan will typically use the LOGNAME environment variable to determine a
userid.
301 302
====

303 304 305
As in any Darshan installation, the darshan-mk-log-dirs.pl script can then be 
used to create the appropriate directory hierarchy for storing Darshan log 
files in the --with-log-path directory.
Philip Carns's avatar
Philip Carns committed
306

307 308 309 310 311
Note that Darshan is not currently capable of detecting the stripe size
(and therefore the Darshan FILE_ALIGNMENT value) on Lustre file systems.
If a Lustre file system is detected, then Darshan assumes an optimal
file alignment of 1 MiB.

312
==== Enabling Darshan instrumentation 
313

314 315
Darshan will automatically install example software module files in the
following locations (depending on how you specified the --prefix option in
316
the previous section):
Philip Carns's avatar
Philip Carns committed
317

318
----
319
/soft/darshan-2.2.3/share/craype-1.x/modulefiles/darshan
320
/soft/darshan-2.2.3/share/craype-2.x/modulefiles/darshan
321
----
Philip Carns's avatar
Philip Carns committed
322

323 324
Select the one that is appropriate for your Cray programming environment
(see the version number of the craype module in `module list`).
Philip Carns's avatar
Philip Carns committed
325

326 327 328 329 330 331 332 333 334 335 336 337 338
If you are using the Cray Programming Environment version 1.x, then you
must modify the corresponding modulefile before using it.  Please see
the comments at the end of the file and choose an environment variable
method that is appropriate for your system.  If this is not done, then
the compiler may fail to link some applications when the Darshan module
is loaded.

If you are using the Cray Programming Environment version 2.x then you can
likely use the modulefile as is.  Note that it pulls most of its
configuration from the lib/pkgconfig/darshan-runtime.pc file installed with
Darshan.

The modulefile that you select can be copied to a system location, or the
339 340
install location can be added to your local module path with the following
command:
Philip Carns's avatar
Philip Carns committed
341

342
----
343
module use /soft/darshan-2.2.3/share/craype-<VERSION>/modulefiles
344
----
Philip Carns's avatar
Philip Carns committed
345

346 347
From this point, Darshan instrumenation can be enabled for all future
application compilations by running "module load darshan".
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

=== Linux clusters using Intel MPI 

Most Intel MPI installations produce dynamic executables by default.  To
configure Darshan in this environment you can use the following example:

----
./configure --with-mem-align=8 --with-log-path=/darshan-logs --with-jobid-env=PBS_JOBID CC=mpicc
----

.Rationale
[NOTE]
====
There is nothing unusual in this configuration except that you should use
the underlying GNU compilers rather than the Intel ICC compilers to compile
Darshan itself.
====

You can use the `LD_PRELOAD` method described earlier in this document to
instrument executables compiled with the Intel MPI compiler scripts.  This
method has been briefly tested using both GNU and Intel compilers.

.Caveat
[NOTE]
====
Darshan is only known to work with C and C++ executables generated by the
Intel MPI suite.  Darshan will not produce instrumentation for Fortran
executables.  For more details please check this Intel forum discussion:

http://software.intel.com/en-us/forums/showthread.php?t=103447&o=a&s=lr
====

380
=== Linux clusters using MPICH 
381 382 383 384 385 386

Follow the generic instructions provided at the top of this document.  The
only modification is to make sure that the `CC` used for compilation is
based on a GNU compiler.  Once Darshan has been installed, it should be
capable of instrumenting executables built with GNU, Intel, and PGI
compilers.
Philip Carns's avatar
Philip Carns committed
387

388 389 390 391 392 393 394 395 396 397 398 399 400
=== Linux clusters using Open MPI

Follow the generic instructions provided at the top of this document for
compilation, and make sure that the `CC` used for compilation is based on a
GNU compiler.

Open MPI typically produces dynamically linked executables by default, which
means that you should use the `LD_PRELOAD` method to instrument executables
that have been built with Open MPI.  Darshan is only compatible with Open
MPI 1.6.4 and newer.  For more details on why Darshan is not compatible with
older versions of Open MPI, please refer to the following mailing list discussion:

http://www.open-mpi.org/community/lists/devel/2013/01/11907.php
401 402 403 404 405 406 407 408 409 410 411 412 413

== Runtime environment variables

The Darshan library honors the following environment variables to modify
behavior at runtime:

* DARSHAN_DISABLE: disables Darshan instrumentation
* DARSHAN_INTERNAL_TIMING: enables internal instrumentation that will print the time required to startup and shutdown Darshan to stderr at run time.
* DARSHAN_LOGHINTS: specifies the MPI-IO hints to use when storing the Darshan output file.  The format is a semicolon-delimited list of key=value pairs, for example: hint1=value1;hint2=value2
* DARSHAN_DISABLE_TIMING: disables the subset of Darshan instrumentation that gathers timing information
* DARSHAN_MEMALIGN: specifies a value for memory alignment (CP_MEM_ALIGNMENT)
* DARSHAN_JOBID: specifies the name of the environment variable to use for the job identifier, such as PBS_JOBID