HPCToolkit is an integrated suite of tools for measurement and analysis of program performance on computers ranging from multicore desktop systems to the nation's largest supercomputers. HPCToolkit provides accurate measurements of a program's work, resource consumption, and inefficiency, correlates these metrics with the program's source code, works with multilingual, fully optimized binaries, has very low measurement overhead, and scales to large parallel systems. HPCToolkit's measurements provide support for analyzing a program execution cost, inefficiency, and scaling characteristics both within and across nodes of a parallel system.
HPCToolkit is released under the 3-clause BSD license. See the file LICENSE for details.
HPCToolkit is available through Spack, simply run:
$ spack install hpctoolkit [+features...]
$ spack load hpctoolkit [+features...]
See the Spack User's Manual for instructions on how to set up Spack for your system and use it to install packages including HPCToolkit.
Available optional features:
+mpi
: Buildhpcprof-mpi
in addition tohpcprof
. Requires MPI (see next section).+python
: Enable the (experimental) Python unwinder. Requires Python (see next section).+papi
: Enable PAPI metrics. Requires PAPI (see next section).+cuda
: Enable CUDA metrics (-e gpu=nvidia
). Requires CUDA (see next section).+level0
: Enable Level Zero metrics (-e gpu=level0
). Requires Level Zero (see next section).+gtpin
: Enable Level Zero instrumentation metrics (-e gpu=level0,inst
). Requires IGC and GTPin (see next section).+opencl
: Enable OpenCL metrics (-e gpu=opencl
). Requires OpenCL (see next section).+rocm
: Enable ROCm metrics (-e gpu=amd
). Requires ROCm (see next section).
Spack will build optional dependencies from source, however in most cases this is unwanted and unwarranted since applications already use these dependencies. In these cases, Spack can be configured to use these external installations as "external packages" in the packages.yaml
configuration file (usually under $HOME/.spack/packages.yaml
).
For example, the configuration for a system ROCm 5.4.3 installation is:
# .../packages.yaml
packages:
hip:
buildable: false
externals:
- spec: [email protected]
prefix: /opt/rocm-5.4.3
hsa-rocr-dev:
buildable: false
externals:
- spec: [email protected]
prefix: /opt/rocm-5.4.3
roctracer-dev:
buildable: false
externals:
- spec: [email protected]
prefix: /opt/rocm-5.4.3
rocprofiler-dev:
buildable: false
externals:
- spec: [email protected]
prefix: /opt/rocm-5.4.3
For more details see the Spack User's Manual.
HPCToolkit supports building from source using Meson.
Install:
- Meson >=1.3.2 and a working C/C++ compiler (GCC, Clang, etc.)
- Boost >=1.71.0, if using CUDA support this must also be built with
visibility=global
(see #815) make
,awk
, andsed
forhpcstruct
measurement directory support (see #704).- (Optional but highly recommended:) ccache
- (Optional:) Wrapped or additional prerequisites
Then run:
$ meson setup builddir/
$ cd builddir/
$ meson compile # -OR- ninja
$ meson test # -OR- ninja test
Do not meson install
/ninja install
here.
Instead, working versions of the tools themselves are available under the meson devenv
, either as individual commands or as a whole subshell environment:
$ meson devenv hpcrun --version # -OR-
$ meson devenv
[builddir] $ hpcrun --version
Additional configuration arguments can be passed to the initial meson setup
or later meson configure
invocations:
-Dmanpages=(disabled|auto|enabled)
: Generate and build man pages. Requires Docutils.-Dhpcprof_mpi=(disabled|auto|enabled)
: Buildhpcprof-mpi
in addition tohpcprof
. Requires MPI.-Dpython=(disabled|auto|enabled)
: Enable the (experimental) Python unwinder. Requires Python headers.-Dpapi=(disabled|auto|enabled)
: Enable PAPI metrics. Requires PAPI.-Dcuda=(disabled|auto|enabled)
: Enable CUDA metrics (-e gpu=nvidia
). Requires CUDA.-Dlevel0=(disabled|auto|enabled)
: Enable Level Zero metrics (-e gpu=level0
). Requires Level Zero.-Dgtpin=(disabled|auto|enabled)
: Also enable Level Zero instrumentation metrics (-e gpu=level0,inst
). Requires Level Zero, IGC/IGA and GTPin.-Dopencl=(disabled|auto|enabled)
: Enable OpenCL metrics (-e gpu=opencl
). Requires OpenCL headers.-Drocm=(disabled|auto|enabled)
: Enable ROCm metrics (-e gpu=amd
). Requires ROCm.-Dvalgrind_annotations=(false|true)
: Inject annotations for debugging with Valgrind.-Dextended_tests=(disabled|auto|enabled)
: Feature option for that require extra dependencies. Request:- CUDA compiler (e.g.
nvcc
) if-Dcuda
support is available/enabled - HIP compiler (
hipcc
) if-Drocm
support is available/enabled - SYCL-compatible compiler (e.g.
icpx
) if-Dlevel0
support is available/enabled
- CUDA compiler (e.g.
Note that many of the features above require additional optional dependencies when enabled.
For Debian Sid/Ubuntu 23.10:
sudo apt-get install git build-essential ccache ninja-build meson cmake pkg-config python3-venv libboost-all-dev libbz2-dev libtbb-dev libelf-dev libdw-dev libunwind-dev libxerces-c-dev libiberty-dev libyaml-cpp-dev libpfm4-dev libxxhash-dev zlib1g-dev
pipx install 'meson>=1.3.2' # Optional but recommended
For older versions:
sudo apt-get install git build-essential ccache ninja-build cmake pkg-config pipx python3 python3-pip python3-venv libboost-all-dev libbz2-dev libtbb-dev libelf-dev libdw-dev libunwind-dev libxerces-c-dev libiberty-dev libyaml-cpp-dev libpfm4-dev libxxhash-dev zlib1g-dev
pipx install 'meson>=1.3.2'
Note that some dependencies may be built from wraps by default, either because they aren't packaged in Debian/Ubuntu (e.g. Dyninst) or because the version is too old. Additional packages can be installed for optional features:
Package | Feature option | Notes |
---|---|---|
mpi-default-dev |
-Dhpcprof_mpi=enabled |
|
libpapi-dev |
-Dpapi=enabled |
|
nvidia-cuda-toolkit |
-Dcuda=enabled |
|
level-zero-dev |
-Dlevel0=enabled |
|
libigdfcl-dev |
-Dlevel0=enabled |
|
libigc-dev |
-Dgtpin=enabled |
|
opencl-headers |
-Dopencl=enabled |
Optional, available as wrap |
rocm-hip-libraries |
-Drocm=enabled |
For RHEL 8:
sudo dnf install git gcc gcc-c++ ccache ninja-build cmake pkg-config python39 python39-pip boost-devel bzip2-devel tbb-devel elfutils-devel xerces-c-devel binutils-devel yaml-cpp-devel libpfm-devel xxhash-devel zlib-devel
python3.9 -m pip install pipx
pipx install 'meson>=1.3.2'
For RHEL 9:
sudo dnf install git gcc gcc-c++ ccache ninja-build cmake pkg-config python3 pipx boost-devel bzip2-devel tbb-devel elfutils-devel xerces-c-devel binutils-devel yaml-cpp-devel libpfm-devel libunwind-devel xxhash-devel zlib-devel
pipx install 'meson>=1.3.2'
For Fedora:
sudo dnf install git gcc gcc-c++ ccache ninja-build cmake pkg-config python3 meson pipx boost-devel bzip2-devel tbb-devel elfutils-devel xerces-c-devel binutils-devel yaml-cpp-devel libpfm-devel libunwind-devel xxhash-devel zlib-devel
pipx install 'meson>=1.3.2' # Optional but recommended
Note that some dependencies may be built from wraps by default, either because they aren't packaged in Fedora/RHEL (e.g. Xed) or because the version is too old. Additional packages can be installed for optional features:
Package | Feature option | Notes |
---|---|---|
openmpi-devel or mpich-devel or mvapich2-devel |
-Dhpcprof_mpi=enabled |
|
papi-devel |
-Dpapi=enabled |
|
cuda-toolkit |
-Dcuda=enabled |
|
level-zero-devel |
-Dlevel0=enabled |
|
intel-igc-opencl-devel |
-Dlevel0=enabled |
|
opencl-headers |
-Dopencl=enabled |
Optional, available as wrap |
rocm |
-Drocm=enabled |
sudo zypper install git gcc12 gcc12-c++ ccache ninja-build cmake pkg-config python311 python311-pip libboost_atomic-devel libboost_chrono-devel libboost_date_time-devel libboost_filesystem-devel libboost_graph-devel libboost_system-devel libboost_thread-devel libboost_timer-devel libbz2-devel tbb-devel libxerces-c-devel binutils-devel yaml-cpp-devel libpfm-devel xxhash-devel zlib-devel
python3.11 -m pip install pipx
pipx install 'meson>=1.3.2'
Note that some dependencies may be built from wraps by default, either because they aren't packaged in SUSE Leap (e.g. Xed) or because the version is too old. Additional packages can be installed for optional features:
Package | Feature option | Notes |
---|---|---|
openmpi-devel or mpich-devel |
-Dhpcprof_mpi=enabled |
|
papi-devel |
-Dpapi=enabled |
|
cuda-toolkit |
-Dcuda=enabled |
|
level-zero-devel |
-Dlevel0=enabled |
|
intel-igc-opencl-devel |
-Dlevel0=enabled |
|
opencl-headers |
-Dopencl=enabled |
Optional, available as wrap |
rocm |
-Drocm=enabled |
HPCToolkit provides Meson wraps for most core dependencies, by default these wraps act as a fallback if they cannot be found otherwise. The convenient part about this is that building HPCToolkit requires little more than Meson and a C/C++ compiler, any missing dependencies will be downloaded and built as subprojects:
Subprojects
bzip2 : YES
dyninst : YES 347 warnings
elfutils : YES 2 warnings
libiberty : YES (from dyninst)
liblzma : YES (from elfutils)
libpfm : YES
libunwind : YES
opencl-headers: YES
tbb : YES
xed : YES
xerces-c : YES
xxhash : YES
yaml-cpp : YES
zstd : YES (from elfutils)
Note that dependencies that are not required will not use a wrap by default, you may need to enable a feature before a wrap will be used (e.g. -Dopencl=enabled
).
If the internet is not available on the system running meson setup
, you may also need to run meson subprojects download
first on a system with internet to download the required sources.
If wraps are not desired, this fallback behavior can be disabled with meson (setup|configure) --wrap-mode=nofallback
.
With this, all dependencies must be installed or otherwise loaded into the build environment, except for force-fallback dependencies meson (setup|configure) --force-fallback-for=dep1,dep2,...
.
See the official Meson documentation for Meson subprojects and wraps for more configuration options and implementation details.
Multiple Dev Container configurations are provided for common distros and vendor software:
ubuntu20.04
,rhel8
,leap15
,fedora39
: CPU-only configurations for common distroscuda*
: Ubuntu 20.04 with NVIDIA CUDA toolkit installedrocm*
: Ubuntu 20.04 with AMD ROCm installedintel
: Ubuntu 20.04 with Intel OneAPI Base Toolkitbare
: Bare minimum dependencies to build HPCToolkit
It is also possible to use custom versions of dependencies, exposing them to Meson however requires some initial understanding of how Meson finds dependencies.
This section is intended as a "peek behind the curtain" of how Meson understands dependencies, and an abridged version of Meson's dependency()
docs for developers.
Meson fetches all dependencies from the "build environment" aka the "machine" (technically two, the build_machine
and the host_machine
, which are only different when cross-compiling).
Whenever Meson configures (i.e. during meson setup
or automatically during meson compile/install/test/etc.
), queries the "machine" as defined by various environment variables:
- Finding programs (
find_program()
) is done by searching thePATH
, - The C/C++ compilers (
CC
/CXX
), which are also affected byCFLAGS
,CXXFLAGS
,LDFLAGS
,INCLUDE_PATH
,LIBRARY_PATH
,CRAY_*
, ... - Dependencies are found via
pkg-config
, which is affected byPKG_CONFIG_PATH
(and otherPKG_CONFIG_*
variables), - Dependencies are found via CMake, which is affected by
CMAKE_PREFIX_PATH
(and otherCMAKE_*
,*_ROOT
and*_ROOT_DIR
variables), - Dependencies with Meson built-in functionality have their own quirks:
- Boost is found under
BOOST_ROOT
, - The CUDA Toolkit is found under
CUDA_PATH
.
- Boost is found under
These variables are frequently altered by commands designed to load software into a shell environment, for example module un/load
.
Meson assumes the "machine" does not change between meson
invocations, however if it does by running the above commands it is possible Meson's cache will not match reality.
In these cases reconfiguration or later compilation may fail due to the inconsistent state between dependencies.
In some cases this can be fixed by clearing Meson's cache (meson configure --clearcache
), Meson will efficiently avoid rebuilding files if the compile command did not actually change.
If build errors persist it may be necessary to wipe the build directory completely (meson setup --wipe
), it is recommended to use ccache
to accelerate the subsequent rebuild.
If you want to use a custom version of a dependency library, install it first and then adjust one or more of the variables mentioned above or corresponding Meson options:
- If the dependency installs a
pkg-config
file (usuallylib/pkgconfig/*.pc
), append the directory containing these files toPKG_CONFIG_PATH
or-Dpkg_config_path
. - In all other cases, append the install prefix (i.e. the directory with
include/
andlib/
orlib64/
) toCMAKE_PREFIX_PATH
or-Dcmake_prefix_path
. - For select dependencies (i.e. if the dependency doesn't install
pkg-config
or CMake files and there is no matchingshare/cmake-*/Modules/Find*.cmake
script in the CMake installation), you may also opt to configure the compiler instead, by:- Extending
-Dc_args
with appropriate-I
include flags and-Dc_link_args
with-L
link flags, or - Extending
CPATH
withinclude/
directories andLIBRARY_PATH
withlib/
orlib64/
directories.
- Extending
Some of these variables can be persisted as Meson built-in options (e.g. -Dpkg_config_path
) or more generally native files.
For example, a native file to use GCC 11 with extra flags and a custom dependency search flags would look like:
# …/my-env.ini
[binaries]
c = ['ccache', 'gcc-11']
cpp = ['ccache', 'g++-11']
[built-in options]
c_args = ['-fstack-protector']
cpp_args = ['-fhardened']
cmake_prefix_path = ['/…/path/to/custom/install/', '/…/path/to/other/install/']
pkg_config_path = ['/…/path/to/custom/install/lib/pkgconfig', '/…/path/to/other/install/lib/pkgconfig']
And then can be used in a build directory by passing additional arguments to meson setup
:
$ meson setup --native-file …/my-elf-dyn.ini builddir/
See doc/developers/meson.ini
for a more complete template listing the settings available in a native file.
Documentation is available at the HPCToolkit home page:
http://hpctoolkit.org/index.html
In particular, there is a PDF User's Manual at:
http://hpctoolkit.org/manual/HPCToolkit-users-manual.pdf
The best way to contact us for bug reports or questions is via the mailing list hpctoolkit-forum =at= rice.edu. The mailman list is technically members-only, but you don't really need to be a member. Just send mail to this address and then we can white-list you.
If reporting a problem, be sure to include sufficient information about your system, what you tried, what went wrong, etc, to allow us follow your steps.
-
type of system, architecture, linux distribution and version, compiler version, glibc version.
-
hpctoolkit version.
-
if a build failure, then the
config.log
file from the build directory and the output frommake
. -
if a run failure, then what command you ran and how it failed.