Table of Contents
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.
Darshan instruments applications via either compile time wrappers or dynamic library preloading. An application that has been instrumented with Darshan will produce a single log file each time it is executed. This log summarizes the I/O access patterns used by the application.
The darshan-runtime instrumentation has traditionally only supported MPI
applications (specifically, those that call MPI_Init() and MPI_Finalize()),
but, as of version 3.2.0, Darshan also supports instrumentation of non-MPI
applications. Regardless of whether MPI is used, Darshan provides detailed
statistics about POSIX level file accesses made by the application.
In the case of MPI applications, Darshan additionally captures details on MPI-IO
and HDF5 level access, as well as limited information about PnetCDF access.
Note that instrumentation of non-MPI applications is currently only supported
in Darshan’s shared library, which applications must LD_PRELOAD.
Starting in version 3.0.0, Darshan also exposes an API that can be used to develop
and add new instrumentation modules (for other I/O library interfaces or to gather
system-specific data, for instance), as detailed in
this document.
Newly contributed modules include a module for gathering system-specific parameters
for jobs running on BG/Q systems, a module for gathering Lustre striping data for
files on Lustre file systems, and a module for instrumenting stdio (i.e., stream I/O
functions like fopen(), fread(), etc).
Starting in version 3.1.3, Darshan also allows for full tracing of application I/O workloads using the newly developed Darshan eXtended Tracing (DxT) instrumentation module. This module can be selectively enabled at runtime to provide high-fidelity traces of an application’s I/O workload, as opposed to the coarse-grained I/O summary data that Darshan has traditionally provided. Currently, DxT only traces at the POSIX and MPI-IO layers. Initial performance results demonstrate the low overhead of DxT tracing, offering comparable performance to Darshan’s traditional coarse-grained instrumentation methods.
Starting in version 3.4.5, Darshan facilitates real-time collection of comprehensive application I/O workload traces through the newly integrated Darshan LDMS data module, known as the darshanConnector. Leveraging the Lightweight Distributed Metric Service (LDMS) streams API, the darshanConnector collects, transports and/or stores traces of application I/O operations instrumented by Darshan at runtime. This module can only be enabled if the LDMS library is included in the Darshan build process. For more information about LDMS or LDMS streams please refer to the official LDMS documentation.
This document provides generic installation instructions, but "recipes" for several common HPC systems are provided at the end of the document as well.
More information about Darshan can be found at the Darshan web site.
Configure and build example (with MPI support).
tar -xvzf darshan-<version-number>.tar.gz cd darshan-<version-number>/ ./prepare cd darshan-runtime/ ./configure --with-log-path=/darshan-logs --with-jobid-env=PBS_JOBID CC=mpicc make make install
Configure and build example (without MPI support).
tar -xvzf darshan-<version-number>.tar.gz cd darshan-<version-number>/ ./prepare cd darshan-runtime/ ./configure --with-log-path=/darshan-logs --with-jobid-env=PBS_JOBID --without-mpi CC=gcc make make install
Explanation of configure arguments:
--with-mem-align=NUM: 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 (default is 8).
--with-jobid-env=NAME (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.
--with-username-env=NAME: this specifies the environment variable that Darshan
should check to determine the username for a job. If not specified, Darshan
will use internal mechanisms to try to determine the username.
LOGNAME environment variable to determine a username, but this method isn’t always reliable (e.g., on Slurm systems, LOGNAME can be wiped when specifying additional environment variables using the --export option to srun). This configure option allows specification of an additional environment variable to extract a username from (e.g., SLURM_JOB_USER).
--with-log-path=DIR (this, or --with-log-path-by-env, is mandatory): This
specifies the parent directory for the directory tree where Darshan logs will
be placed.
darshan-config --log-path command
--with-log-path-by-env=NAME1,NAME2,...: specifies a comma separated list of
environment variables to check at runtime for log path location before the
one set by --with-log-path=DIR at configure time.
--with-log-hints=hint1=x;hint2=y,...: specifies hints to use when writing
the Darshan log file. See ./configure --help for details.
--with-mod-mem=NUM: specifies the maximum amount of memory (in MiB) that
active Darshan instrumentation modules can collectively consume.
--with-zlib=DIR: specifies an alternate location for the zlib development
header and library.
--without-mpi: disables MPI support when building Darshan - MPI support is
assumed if not specified.
--enable-mmap-logs: enables the use of Darshan’s mmap log file mechanism.
--enable-cuserid: enables use of cuserid() at runtime.
--disable-ld-preload: disables building of the Darshan LD_PRELOAD library
--enable-group-readable-logs: sets Darshan log file permissions to allow
group read access.
--disable-exit-wrapper: disables wrapping of _exit() calls as last ditch
shutdown hook for the Darshan library when used in non-MPI mode.
CC=: specifies the C compiler to use for compilation.
Configure arguments for controlling which Darshan modules to use:
--disable-posix-mod: disables compilation and use of Darshan’s POSIX module
(default=enabled)
--disable-mpiio-mod: disables compilation and usee of Darshan’s MPI-IO
module (default=enabled)
--disable-stdio-mod: disables compilation and use of Darshan’s STDIO module
(default=enabled)
--disable-dxt-mod: disables compilation and use of Darshan’s DXT module
(default=enabled)
--enable-hdf5-mod: enables compilation and use of Darshan’s HDF5 module
(default=disabled)
--with-hdf5=DIR: installation directory for HDF5
--enable-hdf5-mod to enable HDF5 modules, --with-hdf5 is only used to additionally provide an HDF5 install prefix.
--enable-pnetcdf-mod: enables compilation and use of Darshan’s PnetCDF
module (default=disabled)
--with-pnetcdf=DIR: installation directory for PnetCDF
--enable-pnetcdf-mod to enable PnetCDF modules, --with-pnetcdf is only used to additionally provide a PnetCDF install prefix.
--disable-lustre-mod: disables compilation and use of Darshan’s Lustre
module (default=enabled)
--enable-mdhim-mod: enables compilation and use of Darshan’s MDHIM module
(default=disabled)
--enable-ldms-mod: enables compilation and use of Darshan’s LDMS runtime module (default=disabled)
--with-ldms=DIR: installation directory for LDMS
--enable-ldms-mod and --with-ldms=DIR to enable runtime data collection via LDMS.
Once darshan-runtime has been installed, you must prepare a location in which to store the Darshan log files and configure an instrumentation method.
This step can be safely skipped if you configured darshan-runtime using the
--with-log-path-by-env option. A more typical configuration uses a static
directory hierarchy for Darshan log
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
placed. The deepest subdirectories will have sticky permissions to enable
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
Regardless of whether a Darshan installation is using the --with-log-path or
--with-log-path-by-env option, end users can display the path (and/or
environment variables) at any time by running darshan-config --log-path
on the command line.
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:
You can also install Darshan via Spack as an alternative to manual download, compilation, and installation. This may be especially convenient for single-user installs. Darshan is divided into two separate packages for the command line utilities and runtime instrumentation. You can install either or both as follows:
spack install darshan-util spack install darshan-runtime
Darshan will generally compile and install fine using a variety of compilers, but we advise using a gcc compiler in Spack to compile Darshan (regardless of what compiler you will use for your applications) to ensure maximum runtime compatibility.
You can use the spack info darshan-runtime query to view the full list of
variants available for the darshan-runtime Spack package. For example, adding a +slurm to
the command line (spack install darshan-runtime+slurm) will cause Darshan
to be compiled with support for gathering job ID information from the Slurm
scheduler.
The following commands will load the Darshan packages once they have been installed:
spack load -r darshan-util spack load -r darshan-runtime
Note that the spack install of darshan-runtime will use an environment
variable named $DARSHAN_LOG_DIR_PATH to indicate where it should store log
files. This variable is set to the user’s home directory by default when
the package is loaded, but it may be overridden.
On Cray systems, you can also perform an additional step to load a
Cray-specific module file. This will make a module called darshan
available as described later in this document in the Cray platform recipe.
It enables automatic instrumentation when using the standard Cray compiler
wrappers.
module use `spack location -i darshan-runtime`/share/craype-2.x/modulefiles
More specific installation "recipes" are provided later in this document for some platforms. This section of the documentation covers general techniques.
Once Darshan has been installed and a log path has been prepared, the next step is to actually instrument applications. The preferred method is to instrument applications at compile time.
This method is applicable to C, Fortran, and C++ MPI applications (regardless of whether they are static or dynamically linked) and is the most straightforward method to apply transparently system-wide. It works by injecting additional libraries and options into the linker command line to intercept relevant I/O calls.
On Cray platforms you can enable the compile time instrumentation by simply loading the Darshan module. It can then be enabled for all users by placing that module in the default environment. As of Darshan 3.2.0 this will instrument both static and dynamic executables, while in previous versions of Darshan this was only sufficient for static executables. See the Cray installation recipe for more details.
For other general MPICH-based MPI implementations, you can generate Darshan-enabled variants of the standard mpicc/mpicxx/mpif90/mpif77 wrappers using the following commands:
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
The resulting *.darshan wrappers will transparently inject Darshan instrumentation into the link step without any explicit user intervention. They can be renamed and placed in an appropriate PATH to enable automatic instrumentation. This method also works correctly for both static and dynamic executables as of Darshan 3.2.0.
For other systems you can enable compile-time instrumentation by either
manually adding the appropriate link options to your command line or
modifying your default MPI compiler script. The darshan-config command
line tool can be used to display the options that you should use:
# Linker options to use for dynamic linking (default on most platforms) # These arguments should go *before* the MPI libraries in the underlying # linker command line to ensure that Darshan can be activated. It should # also ideally go before other libraries that may issue I/O function calls. darshan-config --dyn-ld-flags # linker options to use for static linking # The first set of arguments should go early in the link command line # (before MPI, while the second set should go at the end of the link command # line darshan-config --pre-ld-flags darshan-config --post-ld-flags
The MPICH MPI implementation supports the specification of a profiling library configuration that can be used to insert Darshan instrumentation without modifying the existing MPI compiler script. You can enable a profiling configuration using environment variables or command line arguments to the compiler scripts:
Example for MPICH 3.1.1 or newer:
export MPICC_PROFILE=$DARSHAN_PREFIX/share/mpi-profile/darshan-cc export MPICXX_PROFILE=$DARSHAN_PREFIX/share/mpi-profile/darshan-cxx export MPIFORT_PROFILE=$DARSHAN_PREFIX/share/mpi-profile/darshan-f
Examples for command line use:
mpicc -profile=$DARSHAN_PREFIX/share/mpi-profile/darshan-c <args> mpicxx -profile=$DARSHAN_PREFIX/share/mpi-profile/darshan-cxx <args> mpif77 -profile=$DARSHAN_PREFIX/share/mpi-profile/darshan-f <args> mpif90 -profile=$DARSHAN_PREFIX/share/mpi-profile/darshan-f <args>
Note that unlike the previously described methods in this section, this method will not automatically adapt to static and dynamic linking options. The example profile configurations show above only support dynamic linking.
Example profile configurations are also provided with a "-static" suffix if you need examples for static linking.
This method is applicable to pre-compiled dynamically linked executables as well as interpreted languages such as Python. You do not need to change your compile options in any way. This method works by injecting instrumentation at runtime. It will not work for statically linked executables.
To use this mechanism, set the LD_PRELOAD environment variable to the full
path to the Darshan shared library. The preferred method of inserting Darshan
instrumentation in this case is to set the LD_PRELOAD variable specifically
for the application of interest. Typically this is possible using
command line arguments offered by the mpirun or mpiexec scripts or by
the job scheduler:
mpiexec -n 4 -env LD_PRELOAD /home/carns/darshan-install/lib/libdarshan.so mpi-io-test
srun -n 4 --export=LD_PRELOAD=/home/carns/darshan-install/lib/libdarshan.so mpi-io-test
For sequential invocations of MPI programs, the following will set LD_PRELOAD for process duration only:
env LD_PRELOAD=/home/carns/darshan-install/lib/libdarshan.so mpi-io-test
Other environments may have other specific options for controlling this behavior. Please check your local site documentation for details.
It is also possible to just export LD_PRELOAD as follows, but it is recommended against doing that to prevent Darshan and MPI symbols from being pulled into unrelated binaries:
export LD_PRELOAD=/home/carns/darshan-install/lib/libdarshan.so
For SGI systems running the MPT environment, it may be necessary to set the MPI_SHEPHERD
environment variable equal to true to avoid deadlock when preloading the Darshan shared
library.
Similar to the process described in the previous section, Darshan relies on the
LD_PRELOAD mechanism for instrumenting dynamically-linked non-MPI applications.
This allows Darshan to instrument dynamically-linked binaries produced by non-MPI
compilers (e.g., gcc or clang), extending Darshan instrumentation to new contexts
(like instrumentation of arbitrary Python programs or instrumenting serial
file transfer utilities like cp and scp).
The only additional step required of Darshan non-MPI users is to also set the DARSHAN_ENABLE_NONMPI environment variable to signal to Darshan that non-MPI instrumentation is requested:
export DARSHAN_ENABLE_NONMPI=1
As described in the previous section, it may be desirable to users to limit the scope of Darshan’s instrumentation by only enabling LD_PRELOAD on the target executable:
env LD_PRELOAD=/home/carns/darshan-install/lib/libdarshan.so io-test
Recall that Darshan instrumentation of non-MPI applications is only possible with dynamically-linked applications.
As of Darshan version 3.2.0, Darshan does not necessarily interfere with other profiling tools (particularly those using the PMPI profiling interface). Darshan itself does not use the PMPI interface, and instead uses dynamic linker symbol interception or --wrap function interception for static executables.
As a rule of thumb most profiling tools should appear in the linker command line before -ldarshan if possible.
Darshan’s DXT module provides full tracing of MPI-IO and POSIX read/write APIs. While the DXT module is able to capture finer-grained details compared to traditional Darshan instrumentation, it may exhibit higher runtime and memory overheads. For this reason, DXT support is disabled by default in Darshan, but users can opt-in to DXT instrumentation at runtime by setting their environment as follows:
export DXT_ENABLE_IO_TRACE=1
DXT will trace each I/O operation to files instrumented by Darshan’s MPI-IO and POSIX modules, using a default memory limit of 2 MiB for each module (DXT_POSIX and DXT_MPIIO). Memory usage and a number of other aspects of DXT tracing can be configured as described in section Configuring Darshan library at runtime.
AutoPerf offers two additional Darshan instrumentation modules that may be enabled for MPI applications.
Users can request Darshan to build the APMPI and APXC modules by passing
--enable-apmpi-mod and --enable-apxc-mod options to configure, respectively. Note that these options can be requested independently (i.e., you can build Darshan with APMPI support but not APXC support, and vice versa).
The only prerequsisite for the APMPI module is that Darsan be configured with a MPI-3 compliant compiler. For APXC, the user must obviously be using a Cray XC system and must make the PAPI interface available to Darshan (i.e., by running module load papi, before building Darshan).
If using the APMPI module, users can additionally specify the --enable-apmpi-coll-sync configure option to force Darshan to synchronize before calling underlying MPI routines and to capture additional timing information on how synchronized processes are. Users should note this option will impose additional overheads, but can be useful to help diagnose whether applications are spending a lot of time synchronizing as part of collective communication calls. For this reason, we do not recommend users setting this particular option for production Darshan deployments.
The AutoPerf instrumentation modules are provided as Git submodules to Darshan’s main repository, so if building Darshan source that has been cloned from Git, it is necessary to first retrieve the AutoPerf submodules by running the following command:
git submodule update --init
To fine tune Darshan library settings (e.g., internal memory usage, instrumentation scope, etc.), Darshan provides a couple of mechanisms:
DARSHAN_CONFIG_PATH environment variable
For settings that are specified via a config file and via an environment variable, the environment settings will take precedence.
Users of facility-provided Darshan installs should be mindful that these
installs could define their own default Darshan config file. In this case,
users should double check that DARSHAN_CONFIG_PATH environment variable
is not already set, and if it is, users should consider copying the
default config file as a starting point before applying their own settings.
The Darshan library honors the following settings to modify behavior at runtime:
Table 1. Darshan library config settings
| environment variable setting | config file setting | description |
|---|---|---|
DARSHAN_DISABLE=1 | N/A | Disables Darshan instrumentation. |
DARSHAN_ENABLE_NONMPI=1 | N/A | Enables Darshan's non-MPI mode, required for applications that do not call MPI_Init and MPI_Finalize. |
DARSHAN_CONFIG_PATH=<path> | N/A | Specifies the path to a Darshan config file to load settings from. |
DARSHAN_DUMP_CONFIG=1 | DUMP_CONFIG | Prints the Darshan configuration to stderr at runtime. |
DARSHAN_DISABLE_SHARED_REDUCTION=1 | DISABLE_SHARED_REDUCTION | Disables the step in Darshan aggregation in which files that were accessed by all ranks are collapsed into a single cumulative file record at rank 0. This option retains more per-process information at the expense of creating larger log files. |
DARSHAN_INTERNAL_TIMING=1 | INTERNAL_TIMING | Enables internal instrumentation that will print the time required to startup and shutdown Darshan to stderr at runtime. |
DARSHAN_MODMEM=<val> | MODMEM <val> | Specifies the amount of memory (in MiB) Darshan instrumentation modules can collectively consume (if not specified, a default 4 MiB quota is used). Overrides any `--with-mod-mem` configure argument. |
DARSHAN_NAMEMEM=<val> | NAMEMEM <val> | Specifies the amount of memory (in MiB) Darshan can consume for storing record names (if not specified, a default 1 MiB quota is used). Overrides any `--with-name-mem` configure argument. |
DARSHAN_MEMALIGN=<val> | MEMALIGN <val> | Specifies a value for system memory alignment. Overrides any `--with-mem-align` configure argument (default is 8 bytes). |
DARSHAN_JOBID=<string> | JOBID <string> | Specifies the name of the environment variable to use for the job identifier, such as PBS_JOBID. Overrides `--with-jobid-env` configure argument. |
DARSHAN_LOGHINTS=<string> | LOGHINTS <string> | 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. Overrides any `--with-log-hints` configure argument. |
DARSHAN_LOGPATH=<path> | LOGPATH <path> | Specifies the path to write Darshan log files to. Note that this directory needs to be formatted using the darshan-mk-log-dirs script. Overrides any `--with-log-path` configure argument. |
DARSHAN_MMAP_LOGPATH=<path> | MMAP_LOGPATH <path> | If Darshan's mmap log file mechanism is enabled, this variable specifies what path the mmap log files should be stored in (if not specified, log files will be stored in `/tmp`). |
DARSHAN_LOGFILE=<path> | N/A | Specifies the path (directory + Darshan log file name) to write the output Darshan log to. This overrides the default Darshan behavior of automatically generating a log file name and adding it to a log file directory formatted using darshan-mk-log-dirs script. |
DARSHAN_MOD_DISABLE=<mod_csv> | MOD_DISABLE <mod_csv> | Specifies a list of comma-separated Darshan module names to disable at runtime. |
DARSHAN_MOD_ENABLE=<mod_csv> | MOD_ENABLE <mod_csv> | Specifies a list of comma-separated Darshan module names to enable at runtime. |
DARSHAN_APP_EXCLUDE=<regex_csv> | APP_EXCLUDE <regex_csv> | Specifies a list of comma-separated regexes that match application names that should not be instrumented. This is useful if Darshan is LD_PRELOADed, in which case logs may be generated for many unintended applications. |
DARSHAN_APP_INCLUDE=<regex_csv> | APP_INCLUDE <regex_csv> | Specifies a list of comma-separated regexes that match application names that should be instrumented. This setting is used to override any APP_INCLUDE rules. |
DARSHAN_RANK_EXCLUDE=<rank_csv> | RANK_EXCLUDE <rank_csv> | Specifies a list of comma-separated ranks (or rank ranges) that should not be instrumented. Rank ranges are formatted like "start:end" (if start or end are not specified, the first or last rank is assumed, respectively). Note that the Darshan library will still run on all processes of an application, this setting just controls whether specific ranks are capturing instrumentation data. |
DARSHAN_RANK_INCLUDE=<rank_csv> | RANK_INCLUDE <rank_csv> | Specifies a list of comma-separated ranks (or rank ranges) that should be instrumented. This setting is used to override any RANK_INCLUDE rules. |
DARSHAN_DXT_SMALL_IO_TRIGGER=<val> | DXT_SMALL_IO_TRIGGER <val> | Specifies a floating point percentage (i.e., ".8" would be 80%) indicating a threshold of small I/O operation accesses (defined as accesses smaller than 10 KiB), with DXT trace data being discarded for files that exhibit a percentage of small I/O operations less than this threshold. |
DARSHAN_DXT_UNALIGNED_IO_TRIGGER=<val> | DXT_UNALIGNED_IO_TRIGGER <val> | Specifies a floating point percentage (i.e., ".8" would be 80%) indicating a threshold of unaligned I/O operation accesses (defined as accesses not aligned to the file alignment value determined by Darshan), with DXT trace data being discarded for files that exhibit a percentage of unaligned I/O operations less than this threshold. |
N/A | MAX_RECORDS <val> <mod_csv> | Specifies the number of records to pre-allocate for each instrumentation module given in a comma-separated list. Most modules default to tracing 1024 file records per-process. |
N/A | NAME_EXCLUDE <regex_csv> <mod_csv> | Specifies a list of comma-separated regexes that match record names that should not be instrumented for instrumentation modules given in a comma-separated module list. |
N/A | NAME_INCLUDE <regex_csv> <mod_csv> | Specifies a list of comma-separated regexes that match record names that should be instrumented for instrumentation modules given in a comma-separated module list. This setting is used to override any NAME_EXCLUDE rules. |
DXT_ENABLE_IO_TRACE=1 | N/A | (DEPRECATED) Setting this environment variable enables the DXT (Darshan eXtended Tracing) modules at runtime for all files instrumented by Darshan. Replaced by MODULE_ENABLE setting. |
DARSHAN_EXCLUDE_DIRS=<path_csv> | N/A | (DEPRECATED) Specifies a list of comma-separated paths that Darshan will not instrument at runtime (in addition to Darshan's default exclusion list). Replaced by NAME_EXCLUDE setting. |
DARSHAN_LDMS_ENABLE= | N/A | Switch to initialize LDMS. If not set, no runtime I/O data will be collected. This only needs to be exported (i.e. setting to a value/string is optional). |
DARSHAN_LDMS_ENABLE_<mod_name>= | N/A | Specifies the module data that will be collected during runtime using LDMS streams API. These only need to be exported (i.e. setting to a value/string is optional). |
MOD_DISABLE/ENABLE,
APP_EXCLUDE/INCLUDE, RANK_EXCLUDE/INCLUDE, NAME_EXCLUDE/INCLUDE,
and MAX_RECORDS) may be repeated multiple times rather than
providing comma-separated values, to ease readability.
regex.h interface — refer to the
manpage
for more details on regex syntax.
An example configuration file with annotations is given below
(note that comments are allowed by prefixing a line with #):
# enable DXT modules, which are off by default MOD_ENABLE DXT_POSIX,DXT_MPIIO # allocate 4096 file records for POSIX and MPI-IO modules # (darshan only allocates 1024 per-module by default) MAX_RECORDS 4096 POSIX,MPI-IO # the '*' specifier can be used to apply settings for all modules # in this case, we want all modules to ignore record names # prefixed with "/home" (i.e., stored in our home directory), # with a superseding inclusion for files with a ".out" suffix) NAME_EXCLUDE ^/home * NAME_INCLUDE .out$ * # bump up Darshan's default memory usage to 8 MiB MODMEM 8 # avoid generating logs for git and ls binaries APP_EXCLUDE git,ls # exclude instrumentation for all ranks first RANK_EXCLUDE 0: # then selectively re-include ranks 0-3 and 12:15 RANK_INCLUDE 0:3 RANK_INCLUDE 12:15 # only retain DXT traces for files that were accessed # using small I/O ops 20+% of the time DXT_SMALL_IO_TRIGGER .2
This configuration could be similarly set using environment variables,
though note that both MAX_RECORDS and NAME_EXCLUDE/INCLUDE
settings do not have environment variable counterparts:
export DARSHAN_MOD_ENABLE="DXT_POSIX,DXT_MPIIO" export DARSHAN_MODMEM=8 export DARSHAN_APP_EXCLUDE="git,ls" export DARSHAN_RANK_EXCLUDE="0:" export DARSHAN_RANK_INCLUDE="0:3,12:15" export DARSHAN_DXT_SMALL_IO_TRIGGER=.2
The following recipes provide examples for prominent HPC systems. These are intended to be used as a starting point. You will most likely have to adjust paths and options to reflect the specifics of your system.
This section describes how to compile and install Darshan, as well as how to use a software module to enable and disable Darshan instrumentation on Cray systems.
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 compiler is recommended because it will produce a Darshan library that is interoperable with the widest range of compilers and linkers. On most Cray systems you can enable the GNU programming environment with a command similar to "module swap PrgEnv-intel PrgEnv-gnu". Please see your site documentation for information about how to switch programming environments.
The following example shows how to configure and build Darshan on a Cray system using the GNU programming environment. Adjust the --with-log-path and --prefix arguments to point to the desired log file path and installation path, respectively.
module swap PrgEnv-pgi PrgEnv-gnu ./configure \ --with-log-path=/shared-file-system/darshan-logs \ --prefix=/soft/darshan-3.3.0 \ --with-jobid-env=SLURM_JOBID \ --with-username-env=SLURM_JOB_USER \ CC=cc make install module swap PrgEnv-gnu PrgEnv-pgi
The job ID is set to SLURM_JOBID for use with a Slurm based scheduler.
An additional environment variable for querying a job’s username (SLURM_JOB_USER)
is provided as a fallback in case the default environment variable LOGNAME
is not properly set (e.g., as is the case when using Slurm’s --export option to srun).
The CC variable is configured to point the standard MPI compiler.
If instrumentation of the HDF5 library is desired, additionally load an
acceptable HDF5 module (e.g., module load cray-hdf5-parallel) prior to
building and use the the --enable-hdf5-mod configure argument.
We additionally recommend that you modify Darshan’s generated Cray software
module to include a dependency on the HDF5 software module used — this is
necessary to ensure Darshan library dependencies are satisfied at
application link and run time.
prereq cray-hdf5-parallel
Note that the Darshan-enabled Cray compiler wrappers will always prefer user-supplied HDF5 libraries over the library used to build Darshan. However, due to ABI changes in the HDF5 library, the two HDF5 libraries used must be compatible. Specifically, the HDF5 library versions need to be either both greater than or equal to 1.10 or both less than 1.10. If users use an HDF5 version that is incompatible with Darshan, either link or runtime errors will occur and the user will have to switch HDF5 versions or unload the Darshan module.
Darshan’s default mechanism (clock_gettime()) for retrieving timing
information may introduce more overhead than expected for statically linked
executables on some platforms. The Theta system at the ALCF (as of July
2021) is a notable example. It uses static linking by default (which
prevents the use of the standard vDSO optimization for clock_gettime()
calls), and it’s CPU architecture exhibits relatively high system call
overhead. For Theta and other similar platforms you can explicitly request
that Darshan use the RDTSCP instruction in place of clock_gettime() for
timing purposes. RDTSCP is a non-portable, Intel-specific instruction.
It must be enabled explicitly at configure time, and the base clock
frequency of the compute node CPU must be specified.
This mechanism can be activated on Theta by adding the
--enable-rdtscp=1300000000 to the configure command line (the KNL CPUs on
Theta have a base frequency of 1.3 GHz).
Note that timer overhead is unlikely to be a factor in overall performance unless the application has an edge case workload with frequent sequential I/O operations, such as small I/O accesses to cached data on a single process.
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.
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.
Darshan will automatically install example software module files in the following locations (depending on how you specified the --prefix option in the previous section):
/soft/darshan-3.3.0/share/craype-1.x/modulefiles/darshan /soft/darshan-3.3.0/share/craype-2.x/modulefiles/darshan
Select the one that is appropriate for your Cray programming environment
(see the version number of the craype module in module list).
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 install location can be added to your local module path with the following command:
module use /soft/darshan-3.3.0/share/craype-<VERSION>/modulefiles
From this point, Darshan instrumentation can be enabled for all future application compilations by running "module load darshan".
Most MPICH installations produce dynamic executables by default. To configure Darshan in this environment you can use the following example. We recommend using mpicc with GNU compilers to compile Darshan.
./configure --with-log-path=/darshan-logs --with-jobid-env=PBS_JOBID CC=mpicc
The darshan-gen-* scripts described earlier in this document can be used to create variants of the standard mpicc/mpicxx/mpif77/mpif90 scripts that are Darshan enabled. These scripts will work correctly for both dynamic and statically linked executables.
Most Intel MPI installations produce dynamic executables by default. To configure Darshan in this environment you can use the following example:
./configure --with-log-path=/darshan-logs --with-jobid-env=PBS_JOBID CC=mpicc
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 enable Darshan instrumentation at compile time by adding
darshan-config --dyn-ld-flags options to your linker command line.
Alternatively you can use LD_PRELOAD runtime instrumentation method to
instrument executables that have already been compiled.
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.
You can enable Darshan instrumentation at compile time by adding
darshan-config --dyn-ld-flags options to your linker command line.
Alternatively you can use LD_PRELOAD runtime instrumentation method to
instrument executables that have already been compiled.
In cases where Darshan is not generating a log file for an application, some common things to check are:
darshan-config --log-path command.
For statically linked executables:
nm on it:
> nm test | grep darshan 0000000000772260 b darshan_core 0000000000404440 t darshan_core_cleanup 00000000004049b0 T darshan_core_initialize 000000000076b660 d darshan_core_mutex 00000000004070a0 T darshan_core_register_module
For dynamically linked executables:
> ldd mpi-io-test
linux-vdso.so.1 (0x00007ffd83925000)
libdarshan.so => /home/carns/working/install/lib/libdarshan.so (0x00007f0f4a7a6000)
libmpi.so.12 => /home/carns/working/src/spack/opt/spack/linux-ubuntu19.10-skylake/gcc-9.2.1/mpich-3.3.2-h3dybprufq7i5kt4hcyfoyihnrnbaogk/lib/libmpi.so.12 (0x00007f0f4a44f000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f0f4a241000)
...General:
Ensure that the linker is correctly linking in Darshan’s runtime libraries:
A common mistake is to explicitly link in the underlying MPI libraries (e.g., -lmpich or -lmpichf90)
in the link command, which can interfere with Darshan’s instrumentation
mpicc comipler’s -show flag can be used to examine the invoked link command, for instance
-y option can be used to verify that Darshan is properly intercepting MPI_Init
function (e.g. by setting CFLAGS='-Wl,-yMPI_Init'), which it uses to initialize its runtime structures
/usr/common/software/darshan/3.0.0-pre3/lib/libdarshan.a(darshan-core-init-finalize.o): definition of MPI_Init