HiOp is an optimization solver for solving certain mathematical optimization problems expressed as nonlinear programming problems. HiOp is a lightweight HPC solver that leverages application's existing data parallelism to parallelize the optimization iterations by using specialized parallel linear algebra kernels.
Please cite the user manual whenever HiOp is used:
@TECHREPORT{hiop_techrep,
title={{HiOp} -- {U}ser {G}uide},
author={Petra, Cosmin G. and Chiang, NaiYuan and Jingyi Wang},
year={2018},
institution = {Center for Applied Scientific Computing, Lawrence Livermore National Laboratory},
number = {LLNL-SM-743591}
}
In addition, when using the quasi-Newton solver please cite:
@ARTICLE{Petra_18_hiopdecomp,
title = {A memory-distributed quasi-Newton solver for nonlinear programming problems with a small number of general constraints},
journal = {Journal of Parallel and Distributed Computing},
volume = {133},
pages = {337-348},
year = {2019},
issn = {0743-7315},
doi = {https://doi.org/10.1016/j.jpdc.2018.10.009},
url = {https://www.sciencedirect.com/science/article/pii/S0743731518307731},
author = {Cosmin G. Petra},
}
and when using the the PriDec solver please cite:
@article{wang2023,
author = {J. Wang and C. G. Petra},
title = {A Sequential Quadratic Programming Algorithm for Nonsmooth Problems with Upper-$\mathcal{C}^2$ Objective},
journal = {SIAM Journal on Optimization},
volume = {33},
number = {3},
pages = {2379-2405},
year = {2023},
doi = {10.1137/22M1490995}
}
@INPROCEEDINGS{wang2021,
author={J. Wang and N. Chiang and C. G. Petra},
booktitle={2021 20th International Symposium on Parallel and Distributed Computing (ISPDC)},
title={An asynchronous distributed-memory optimization solver for two-stage stochastic programming problems},
year={2021},
volume={},
number={},
pages={33-40},
doi={10.1109/ISPDC52870.2021.9521613}}
}
HiOp uses a CMake-based build system. A standard build can be done by invoking in the 'build' directory the following
$> cmake ..
$> make
$> make test
$> make installThis sequence will build HiOp, run integrity and correctness tests, and install the headers and the library in the directory '_dist-default-build' in HiOp's root directory.
Command make test runs extensive tests of the various modules of HiOp to check integrity and correctness. The tests suite range from unit testing to solving concrete optimization problems and checking the performance of HiOp solvers on these problems against known solutions. By default make test runs mpirun locally, which may not work on some HPC machines. For these HiOp allows using bsub to schedule make test on the compute nodes; to enable this, the use should use -DHIOP_TEST_WITH_BSUB=ON with cmake when building and run make test in a bsub shell session, for example,
bsub -P your_proj_name -nnodes 1 -W 30
make test
CTRL+D
The installation can be customized using the standard CMake options. For example, one can provide an alternative installation directory for HiOp by using
$> cmake -DCMAKE_INSTALL_PREFIX=/usr/lib/hiop ..'- Enable/disable MPI: -DHIOP_USE_MPI=[ON/OFF] (by default ON)
- GPU support: -DHIOP_USE_GPU=ON. MPI can be either off or on. For more build system options related to GPUs, see "Dependencies" section below.
- Enable/disable "developer mode" build that enforces more restrictive compiler rules and guidelines: -DHIOP_DEVELOPER_MODE=ON. This option is by default off.
- Additional checks and self-diagnostics inside HiOp meant to detect abnormalities and help to detect bugs and/or troubleshoot problematic instances: -DHIOP_DEEPCHECKS=[ON/OFF] (by default ON). Disabling HIOP_DEEPCHECKS usually provides 30-40% execution speedup in HiOp. For full strength, it is recommended to use HIOP_DEEPCHECKS with debug builds. With non-debug builds, in particular the ones that disable the assert macro, HIOP_DEEPCHECKS does not perform all checks and, thus, may overlook potential issues.
For example:
$> cmake -DHIOP_USE_MPI=ON -DHIOP_DEEPCHECKS=ON ..
$> make
$> make test
$> make install- -DCMAKE_BUILD_TYPE=Release will build the code with the optimization flags on
- -DCMAKE_CXX_FLAGS="-O3" will enable a high level of compiler code optimization
A complete list of dependencies is maintained here.
For a minimal build, HiOp requires LAPACK and BLAS. These dependencies are automatically detected by the build system. MPI is optional and by default enabled. To disable use cmake option '-DHIOP_USE_MPI=OFF'.
Please note that HiOp has support for GPU computations, for both NVIDIA and AMD hardware, via RAJA and Umpire performance portability libraries. To quickly enable the use of GPUs, use cmake with '-DHIOP_USE_GPU=ON'.
If you would like to use our public package.py:
spack install hiop@developAdditional information about using Spack can be found at spack.readthedocs.io or within HiOp's package.py.
HiOp supports three types of optimization problems, each with a separate input formats in the form of the C++ interfaces hiopInterfaceDenseConstraints,hiopInterfaceSparse and hiopInterfaceMDS. These interfaces are specified in hiopInterface.hpp and documented and discussed as well in the user manual.
hiopInterfaceDenseConstraints interface supports NLPs with billions of variables with and without bounds but only limited number (<100) of general, equality and inequality constraints. The underlying algorithm is a limited-memory quasi-Newton interior-point method and generally scales well computationally (but it may not algorithmically) on thousands of cores. This interface uses MPI for parallelization
hiopInterfaceSparse interface supports general sparse and large-scale NLPs. This functionality is similar to that of the state-of-the-art Ipopt (without being as robust and flexible as Ipopt is). Acceleration for this class of problems can be achieved via OpenMP or CUDA, however, this is work in progress and you are encouraged to contact HiOp's developers for up-to-date information.
hiopInterfaceMDS interface supports mixed dense-sparse NLPs and achives parallelization using GPUs and RAJA portability abstraction layer.
More information on the HiOp interfaces are here.
Users are highly encouraged to report any issues they found when using HiOp.
HiOp has been developed under the financial support of:
- Department of Energy, Office of Advanced Scientific Computing Research (ASCR): Exascale Computing Program (ECP) and Applied Math Program.
- Department of Energy, Advanced Research Projects Agency-Energy (ARPA‑E)
- Lawrence Livermore National Laboratory Institutional Scientific Capability Portfolio (ISCP)
- Lawrence Livermore National Laboratory, through the LDRD program
HiOp is written by Cosmin G. Petra ([email protected]), Nai-Yuan Chiang ([email protected]), Jingyi "Frank" Wang ([email protected]), and Tucker Hartland ([email protected]) from LLNL and has received important contributions from Asher Mancinelli (PNNL), Slaven Peles (ORNL), Cameron Rutherford (PNNL), Jake K. Ryan (PNNL), and Michel Schanen (ANL).
Copyright (c) 2017-2025, Lawrence Livermore National Security, LLC. All rights reserved. Produced at the Lawrence Livermore National Laboratory. LLNL-CODE-742473. HiOp is free software; you can modify it and/or redistribute it under the terms of the BSD 3-clause license. See COPYRIGHT and LICENSE for complete copyright and license information.
