GPTune is an autotuning framework that relies on multitask and transfer learnings to help solve the underlying black-box optimization problem using Bayesian optimization methodologies.
- GPTune
- Table of Contents
- (High-performance and parallel tuning) GPTune is designed to tune applications running on large-scale cluster systems and can exploit distributed memory parallelism for accelerating surrogate modeling.
- (Multitask learning-based autotuning) GPTune supports multitask learning-based autotuning that allows us to tune multiple tuning problems simultaneously. Multitask learning would predict an optimal tuning parameter configuration using a fewer number of evaluations compared to single task autotuning by modeling the linear dependence of the multiple tasks.
- (Transfer learning-based autotuning) GPTune supports transfer learning-based autotuning to tune the given tuning task by leveraging already available performance data collected for different tasks. Different tasks can mean different input problem sizes or the same input problem on different machine and/or software settings.
- (GPTuneBand (multi-fidelity autotuning)) Multi-fidelity tuning uses multiple fidelity levels to guide sampling (generating many cheap samples from lower-fidelity levels). GPTuneBand combines multitask learning with a multi-armed bandit strategy to guide sampling of the given tuning problem.
- (Multi-objective tuning) Beyond the classical single-objective tuning, GPTune supports multi-objective tuning that uses NSGA2 algorithm to maximize multiple EI functions for multiple objectives. For an objective, users can also specify whether they want to optimize (minimize) the objective within the given range, or they just want the objective is within the given range.
- (Unified interface for different autotuners) GPTune uses a unified Python interface and supports using several different autotuners.
- (History database) We provide an autotuning database called GPTune history database which allows users to save and re-use performance data to reduce the cost of the expensive black-box objective function. The history database enables several useful autotuning capabilities. The details are outlined in the GPTune and history database project webpage at https://gptune.lbl.gov/about.
GPTune is part of the xSDK4ECP effort supported by the Exascale Computing Project (ECP). Our GPTune website at https://gptune.lbl.gov provides a shared database repository where the user can share their tuning performance data with other users.
To best support the various needs of different users, we provide several GPTune installation modes, as detailed below. Users may choose an installation mode in the decision tree depending on their needs.
Figure 1: Installation Decision Tree.
The easiest way to install GPTune is to install the lite version of GPTune, which skips many software dependencies required by advanced GPTune features. As a result, the lite version only supports the basic functionalities and the tuning performance can be compromised. The lite version is what we highly recommend to try out first before you decide to use GPTune for more serious tasks.
GPTune lite version can be installed simply by:
git clone https://github.com/gptune/GPTune
cd GPTune
pip install --user -r requirements.txt
In case some packages (particularly openturns or pygmo) in requirements.txt fail to be installed, try:
export GPTUNE_LITE_MODE=1
pip install --user -r requirements_lite.txt
The following example build scripts are available for a collection of tested systems.
The following script installs everything from scratch and can take up to 2 hours depending on the users' machine specifications. If "MPIFromSource=0", you need to set PATH, LIBRARY_PATH, LD_LIBRARY_PATH and MPI compiler wrappers when prompted.
config_cleanlinux.sh
Note that this would require superuser (sudo) account of the system. If this is not the case, one can use
config_linux.sh
The following script installs everything from scratch and can take up to 2 hours depending on the users' machine specifications. The user may need to set pythonversion, gccversion, openblasversion, lapackversion on the top of the script to the versions supported by your homebrew software.
config_macbook.zsh
The following script installs GPTune with mpi, python, compiler, cudatoolkit and cmake modules on Perlmutter. Note that you need to set "proc=milan #(CPU nodes) or gpu #(GPU nodes)", "mpi=openmpi or craympich" and "compiler=gnu". Setting mpi=craympich will only support the RCI mode.
config_perlmutter.sh
The following script installs GPTune with mpi, python, compiler, cuda and cmake modules on Summit. Note that you can set "proc=power9", "mpi=spectrummpi" and "compiler=gnu". Currently, only the RCI mode can be used on Summit.
config_summit.sh
The following script installs GPTune with the PrgEnv-gnu module on Frontier. Note that you can set "proc=EPYC", "mpi=craympich" and "compiler=gnu". Currently, only the RCI mode can be used on Frontier.
config_frontier.sh
One can also consider using Spack (https://spack.io/). To install and test GPTune using Spack (the develop branch of the spack github repo is highly recommended), one simply needs:
spack install gptune@master
spack load gptune@master
For single-node systems, Nix may be used to install GPTune and all its dependencies on single-node systems, including personal computers and cloud servers (both with and without root access). Nix pulls in independent copies of GPTune's dependencies, and as a result it will neither affect nor be affected by the state of your system's packages.
If you have root access, run this command to automatically install Nix, then immediately proceed to step 2:
sh <(curl -L https://nixos.org/nix/install) --daemon
For more details, see the manual for full details).
If you do not have root access, you can install Nix as an unpriviliged user using one of [these] methods. For systems supporting user namespaces (follow the instructions here to check for user namespace support), including Debian, Ubuntu, and Arch, nix-user-chroot is recommended; the following steps may be used to install it. First, download the appropriate static binary for your hardware platform:
#replace the link below with the appropriate build for your architecture
wget -O nix-user-chroot https://github.com/nix-community/nix-user-chroot/releases/download/1.2.2/nix-user-chroot-bin-1.2.2-x86_64-unknown-linux-musl
chmod +x nix-user-chroot
#optionally, add nix-user-chroot to $PATH - you'll be running it a lot
Then, select an installation location. For this example, we will use ~/.nix
. Note that Nix will perform a significant amount of disk I/O to this location, so make sure that this directory is not located on a network drive (NFS, etc.) or builds may be slowed by up to an order of magnitude (for example, we recommend that UC Berkeley Millennium cluster users use a folder in /scratch
or /nscratch
instead). You may then install nix with:
mkdir -m 0755 ~/.nix
./nix-user-chroot ~/.nix bash -c "curl -L https://nixos.org/nix/install | bash"
You may now enter the nix chroot environment with
./nix-user-chroot ~/.nix bash -l
This works much like a python virtualenv or conda shell - it will drop you into a environment where the Nix package manager and tools you have installed with Nix are available. As with a python virtualenv, you must be inside this environment in order to access tools (e.g. GPTune) that are installed with Nix (i.e. you must run it in each shell where you need these tools). All programs, files, etc. outside the environment should be accessible from within the environment as well.
Troubleshooting note: if you run into "out of space" errors during builds, set the TMPDIR
environment variable when you run this command to a location on a disk with plenty of space, e.g. TMPDIR=/scratch/dinh/tmp nix-user-chroot /scratch/dinh/.nix bash
If you do not have root access and your system does not support user namespaces, you can install Nix using proot.
Nix flakes, which we use to build GPTune, are technically an experimental feature in nix (this is not a mark of instability - flakes have existed and been widely used for years, but remain marked as experimental since there's a slim possibility that their interface might change). As a result, they must be manually enabled, which can be done by adding this line:
experimental-features = nix-command flakes
to any one (or more) of the following locations (if the file(s) in question don't exist, feel free to create them):
~/.config/nix/nix.conf
(recommended, affects your user account only)/etc/nix/nix.conf
(if you have root access, and want to enable flakes for everyone on your system)/nix/etc/nix/nix.conf
(for chroot-based installs only)
Clone the GPTune repo and cd into its directory:
git clone https://github.com/gptune/GPTune
cd GPTune
then run
nix develop
to enter an environment where the python
executable has all the dependencies needed.
Alternatively, if you just want the C++ libraries for GPTune (e.g. to link with), run nix build .#gptune-libs
, which will put the librarires in result/gptune
.
GPTune full version relies on OpenMPI (4.0 or higher) or CrayMPICH (8.1.23 or higher), Python (3.7 or higher), BLAS/LAPACK, SCALAPACK (2.1.0 or higher) and mpi4py, which need to be installed by the user. In what follows, we assume Python, BLAS/LAPACK have been installed (with the same compiler version):
git clone https://github.com/gptune/GPTune
cd GPTune # this is the GPTUNEROOT directory
export MPICC=path-to-c-compiler-wrapper # see next subsection to install OpenMPI from source if one doesn't yet have one installed.
export MPICXX=path-to-cxx-compiler-wrapper
export MPIF90=path-to-f90-compiler-wrapper
export MPIRUN=path-to-mpirun
export BLAS_LIB=path-to-blas-lib
export LAPACK_LIB=path-to-lapack-lib
export GPTUNEROOT=path-to-gptune-root-directory
export SITE_PACKAGES_PATH=path-to-your-site-packages
The rest can be installed as follows:
cd $GPTUNEROOT
wget https://download.open-mpi.org/release/open-mpi/v4.0/openmpi-4.0.1.tar.bz2
bzip2 -d openmpi-4.0.1.tar.bz2
tar -xvf openmpi-4.0.1.tar
cd openmpi-4.0.1/
./configure --prefix=$PWD --enable-mpi-interface-warning --enable-shared --enable-static --enable-cxx-exceptions CC=$MPICC CXX=$MPICXX F77=$MPIF90 FC=$MPIF90 --enable-mpi1-compatibility --disable-dlopen
make -j4
make install
export PATH=$PATH:$GPTUNEROOT/openmpi-4.0.1/bin
export MPICC="$GPTUNEROOT/openmpi-4.0.1/bin/mpicc"
export MPICXX="$GPTUNEROOT/openmpi-4.0.1/bin/mpicxx"
export MPIF90="$GPTUNEROOT/openmpi-4.0.1/bin/mpif90"
export LD_LIBRARY_PATH=$GPTUNEROOT/openmpi-4.0.1/lib:$LD_LIBRARY_PATH
export LIBRARY_PATH=$GPTUNEROOT/openmpi-4.0.1/lib:$LIBRARY_PATH
cd $GPTUNEROOT
wget http://www.netlib.org/scalapack/scalapack-2.1.0.tgz
tar -xf scalapack-2.1.0.tgz
cd scalapack-2.1.0
rm -rf build
mkdir -p build
cd build
mkdir -p install
cmake .. \
-DBUILD_SHARED_LIBS=ON \
-DCMAKE_C_COMPILER=$MPICC \
-DCMAKE_Fortran_COMPILER=$MPIF90 \
-DCMAKE_INSTALL_PREFIX=./install \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-DCMAKE_Fortran_FLAGS="-fopenmp" \
-DBLAS_LIBRARIES="$BLAS_LIB" \
-DLAPACK_LIBRARIES="$LAPACK_LIB"
make
make install
export SCALAPACK_LIB="$PWD/install/lib/libscalapack.so"
cd $GPTUNEROOT
git clone https://github.com/mpi4py/mpi4py.git
cd mpi4py/
python setup.py build --mpicc="$MPICC -shared"
python setup.py install --user
export PYTHONPATH=$PYTHONPATH:$PWD
GPTune also depends on several external Python libraries as listed in the requirements.txt
file, including numpy, joblib, scikit-learn, scipy, statsmodels, pyaml, matplotlib, GPy, openturns,lhsmdu, ipyparallel, opentuner, hpbandster, pygmo, filelock, requests, pymoo and cloudpickle. These Python libraries can all be installed through the standard Python repository through the pip tool.
cd $GPTUNEROOT
env CC=$MPICC pip install --user -r requirements.txt
In addition, the following MPI-enabled component needs to be installed:
cd $GPTUNEROOT
export PYTHONPATH=$PYTHONPATH:$PWD
mkdir -p build
cd build
cmake .. \
-DBUILD_SHARED_LIBS=ON \
-DCMAKE_CXX_COMPILER=$MPICXX \
-DCMAKE_C_COMPILER=$MPICC \
-DCMAKE_Fortran_COMPILER=$MPIF90 \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-DGPTUNE_INSTALL_PATH=$SITE_PACKAGES_PATH \
-DTPL_BLAS_LIBRARIES="$BLAS_LIB" \
-DTPL_LAPACK_LIBRARIES="$LAPACK_LIB" \
-DTPL_SCALAPACK_LIBRARIES=$SCALAPACK_LIB
make
One can also try the prebuilt docker image of GPTune to test its functionality
docker pull liuyangzhuan/gptune:4.5
docker run -it -v $HOME:$HOME liuyangzhuan/gptune:4.5
There are a few examples included in GPTune, each example is located in a seperate directory ./examples/[application_name]. The user needs to edit examples/[application_name]/.gptune/meta.json to define machine information and software dependency, before running the tuning examples.
Please take a look at the following two scripts to run the complete examples. Note that these scripts first load ./run_env.sh (the user needs to modify this file to define appropriate runtime variables, machine and software information), generate the examples/[application_name]/.gptune/meta.json file, and then invoke the tuning experiment.
https://github.com/gptune/GPTune/blob/master/examples/GPTune-Demo/run_examples.sh
https://github.com/gptune/GPTune/blob/master/run_ppopp.sh
The file demo.py
in the examples/GPTune-Demo
folder shows how to describe the autotuning problem for a sequential objective function and how to invoke GPTune
cd $GPTUNEROOT/examples/GPTune-Demo
edit .gptune/meta.json
python ./demo.py
or
edit ../../run_env.sh
bash run_examples.sh
The files scalapack_*.py
in the examples/Scalapack-PDGEQRF
folder shows how to tune the parallel QR factorization subroutine PDGEQRF with different features of GPTune.
cd $GPTUNEROOT/examples/Scalapack-PDGEQRF
edit .gptune/meta.json
$MPIRUN -n 1 python ./scalapack_MLA.py -mmax 1000 -nmax 1000 -nprocmin_pernode 1 -ntask 2 -nrun 20 -optimization 'GPTune'
or
edit ../../run_env.sh
bash run_examples.sh
First, SuperLU_DIST needs to be installed with the same OpenMPI/CrayMPICH and BLAS/LAPACK as the above.
cd $GPTUNEROOT/examples
git clone https://github.com/xiaoyeli/superlu_dist.git
cd superlu_dist
wget http://glaros.dtc.umn.edu/gkhome/fetch/sw/parmetis/parmetis-4.0.3.tar.gz
tar -xf parmetis-4.0.3.tar.gz
cd parmetis-4.0.3/
mkdir -p install
make config shared=1 cc=$MPICC cxx=$MPICXX prefix=$PWD/install
make install
cd ../
export PARMETIS_INCLUDE_DIRS="$PWD/parmetis-4.0.3/metis/include;$PWD/parmetis-4.0.3/install/include"
export PARMETIS_LIBRARIES=$PWD/parmetis-4.0.3/install/lib/libparmetis.so
mkdir -p build
cd build
cmake .. \
-DCMAKE_CXX_FLAGS="-Ofast -std=c++11 -DAdd_ -DRELEASE" \
-DCMAKE_C_FLAGS="-std=c11 -DPRNTlevel=0 -DPROFlevel=0 -DDEBUGlevel=0" \
-DBUILD_SHARED_LIBS=OFF \
-DCMAKE_CXX_COMPILER=$MPICXX \
-DCMAKE_C_COMPILER=$MPICC \
-DCMAKE_Fortran_COMPILER=$MPIF90 \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-DTPL_BLAS_LIBRARIES="$BLAS_LIB" \
-DTPL_LAPACK_LIBRARIES="$LAPACK_LIB" \
-DTPL_PARMETIS_INCLUDE_DIRS=$PARMETIS_INCLUDE_DIRS \
-DTPL_PARMETIS_LIBRARIES=$PARMETIS_LIBRARIES
make pddrive_spawn
Note that pddrive_spawn
is a modified application driver that will be launched by GPTune via MPI spawning (see the Usage section). The files superlu_*.py
in the examples/SuperLU_DIST
folder shows how to tune the performance of sparse LU factorization with different features of GPTune.
cd $GPTUNEROOT/examples/SuperLU_DIST
edit .gptune/meta.json
$MPIRUN -n 1 python ./superlu_MLA_MO.py -nprocmin_pernode 1 -ntask 1 -nrun 10 -optimization 'GPTune'
or
edit ../../run_env.sh
bash run_examples.sh
In order to autotune a certain application, three spaces have to be defined through an instance of the Space class.
- Input Space (IS): this space defines the problems that the application targets. Every point in this space represents one instance of a problem. In the context of GPTune, the word task means application problem.
- Parameter Space (PS): this space defines the application parameters to be tuned. A point in this space represents a combination of the parameters. The goal of the tuner is to find the best possible combination that minimizes the objective function of the application.
- Output Space (OS): this space defines the result(s) of the application, i.e., the objective of the application to be optimized. For examples, this can be runtime, memory or energy consumption in HPC applications or prediction accuracy in machine learning applications.
Every dimension of the above mentioned spaces is defined by a Parameter object. Every parameter is defined by its name, type and range or set of values. Three types of parameters can be defined:
- Real: defines floating point parameters. The range of values that the parameter spans should be defined in the range argument.
- Integer: defines integer parameters. The range of values that the parameter spans should be defined in the range argument.
- Categorical: defines parameters that take their values in a set or list of (string) values. The list of valid values defining the parameter should be defined in the values argument. (Note: If the problems the application targets cannot be defined in a Cartesian space, the user can simply give a list of problems (as a Categorical parameter) in the definition of the task space.)
Not all points in the task or input spaces correspond to valid problems or parameter configurations. Constraints might exist that define the validity of a given combination of input parameters and problem description parameters results. Two ways exist to define constraints in GPTune:
- Strings: the user can define a Python statement in a string. The evaluation of that statement should be a boolean.
- Functions: the user can define a Python function that returns a boolean.
The user need to define a Python function representing the objective function to be optimized
def objectives(point)
# extract the parameters from 'point', invoke the application code, and store the objective function value in 'result'
return result
Here point is a dictionary containing key pairs representing task and tuning parameters. If the application code is distributed-memory parallel, the user needs to modify the application code (in C, C++, Fortran, Python, etc.) with the MPI spawning syntax and invoke the code using MPI.COMM_SELF.Spawn from mpi4py. Please refer to the user guide for details https://github.com/gptune/GPTune/blob/master/Doc/GPTune_UsersGuide.pdf.
The user having additional knowledge about the application can help speed up or improve the result of the tuning process by passing a performance model(s) of the objective function to be optimized.
These models are defined through Python functions following similarly to the constraints definition.
Once the parameters and spaces (and optionally constraints and models) are defined, an object of the GPTune class has to be instantiated. Then, the different kinds of tuning techniques (MLA, ...) can be called through it.
[Manual] GPTune UsersGuide
[Manual] GPTune History Database & Shared Repository
[Tutorial] GPTune: Performance Autotuner for ECP Applications, Tutorial at ECP Annual Meeting, April 14, 2021
[Talk] GPTune: Multitask Learning for Autotuning Exascale Applications, PPoPP 2021, February 27, 2021
[Talk] Autotuning exascale applications with Gaussain Process Regression, E-NLA Seminar, October 14, 2020
Y. Liu, W.M. Sid-Lakhdar, O. Marques, X. Zhu, C. Meng, J.W. Demmel, and X.S. Li. "GPTune: multitask learning for autotuning exascale applications", in Proceedings of the 26th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP '21). Association for Computing Machinery, New York, NY, USA, 234-246. DOI:https://doi.org/10.1145/3437801.3441621
@inproceedings{liu2021gptune,
title={GPTune: multitask learning for autotuning exascale applications},
author={Liu, Yang and Sid-Lakhdar, Wissam M and Marques, Osni and Zhu, Xinran and Meng, Chang and Demmel, James W and Li, Xiaoye S},
booktitle={Proceedings of the 26th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming},
pages={234--246},
year={2021}
}
There are multiple extensions/variants of GPTune tuning techniques, extending the techniques in the first GPTune PPoPP 2021 paper. They are incorporated into this GPTune package or callable from the GPTune interface.
Xinran Zhu, Yang Liu, Pieter Ghysels, David Bindel, and Xiaoye S. Li. "GPTuneBand: Multi-task and Multi-fidelity Autotuning for Large-scale High Performance Computing Applications", in SIAM Conference on Parallel Processing for Scientific Computing, 2022.
Younghyun Cho, James W. Demmel, Jacob King, Xiaoye S. Li, Yang Liu, and Hengrui Luo. "Harnessing the Crowd for Autotuning High-Performance Computing Applications", in IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2023.
Hengrui Luo, Younghyun Cho, James W Demmel, Xiaoye S Li, Yang Liu. "Hybrid Parameter Search and Dynamic Model Selection for Mixed-Variable Bayesian Optimization", in Journal of Computational and Graphical Statistics, 2024.
Hengrui Luo, Younghyun Cho, James W Demmel, Igor Kozachenko, Xiaoye S Li, and Yang Liu. "Non-smooth Bayesian optimization in tuning scientific applications", The International Journal of High Performance Computing Applications, 2024.
GPTune Copyright (c) 2019, The Regents of the University of California, through Lawrence Berkeley National Laboratory (subject to receipt of any required approvals from the U.S.Dept. of Energy) and the University of California, Berkeley. All rights reserved.
If you have questions about your rights to use or distribute this software, please contact Berkeley Lab's Intellectual Property Office at [email protected].
NOTICE. This Software was developed under funding from the U.S. Department of Energy and the U.S. Government consequently retains certain rights. As such, the U.S. Government has been granted for itself and others acting on its behalf a paid-up, nonexclusive, irrevocable, worldwide license in the Software to reproduce, distribute copies to the public, prepare derivative works, and perform publicly and display publicly, and to permit other to do so.