The Optimization Plugin Package (aka “op”)

op is a lightweight general optimization solver interface. The primary purpose of op is to simplify the process of integrating different optimization solvers (serial or parallel) with scalable parallel physics engines. By design, it has several features that help make this a reality.

  • The core abstraction interface was developed to encompass a large class of optimization problems in an optimizer agnostic way. This allows us to describe the optimization problem once and then use a variety of supported op optimizers with none to little code-changes.
  • The op core abstract interface is made up of lightweight wrappers that make it easy to integrate with existing simulation codes. This makes integration less intrusive and helps maintain the “integrity” of the physics code. An added benefit is that we can “re-run” an optimization problem on different physics engines by making only a few changes in op.
  • The op interface includes an assortment of utility methods to perform complicated Halo-exchange patterns based on very little information as well as converting optimization primitives from one framework (e.g. nlopt to ipopt). In addition, the op::utility and op::mpi interface is available for users to customize their own MPI patterns as necessary.
  • Lastly a black-box optimizer plugin interface is provided in op to allow for use of proprietary optimization engines without explicit reference in the source code.

The serial nlopt optimization library is used to demonstrate how op bridges the gap between serial optimizers and parallel simulation codes. op::NLopt will run the optimization problem in serial but the physics, objectives, constraints, and respective gradients are computed in parallel. From a user perspective, optimization problems can be described abstractly and op guaranatees portability for different optimization engines and optimization problem configurations.

Install

The following will install to ../install.

mkdir build
cd build
cmake -DBLT_SOURCE_DIR=<blt-dir> -C ../host_configs/quartz.cmake -DCMAKE_INSTALL_PREFIX=$PWD/../install ..

Using OP in another project

The following cmake line can be added to CMakeLists.txt to find the required package. OP_DIR must be defined either as a cmake commandline option, in CMakeCache.txt, or in a toolchain file.

find_package(op CONFIG REQUIRED
                PATHS ${OP_DIR}/lib/cmake)

Spack package

There is a spack package for use with uberenv available in <op-root>/spack/packages/op. It can be built with gcc with vanilla spack, but clang support requires some toolchain fixes available in serac.

Documentation

More user documentation is available in sphinx and API documentation in doxygen. To build the documentation type in the following in the build directory:
make docs

Spinx documentation is available in <build-dir>/src/docs/sphinx or <install-dir>/docs/sphinx. Sphinx documentation covers core op abstraction concepts, a suggested migration guide in going from an existing optimization solver to op‘s more general interface, and also documentation for the TwoCnsts Rosenbrock problem. Some documentation for the simplified MPI methods is also provided.

Doxygen API documentation is available in <docs-dir>/sphinx/op_docs/html/doxygen.

The documentation is also available online!

Included op Examples

Currently there are a few examples

  • bin/demo – demo is a program to test plugin loading with a dummy optimizer call test_optimizer.
  • tests/TwoCnsts – A set of tests that demonstrate different ways of using nlopt on a common two constraint rosenbrock problem.
  • VariableMap – A set of tests demonstrating a more complicated halo-exchange pattern and how to both use built-in op patterns or build custom communication patterns using op::mpi and op::utility convenience methods.

tests/TwoCnsts

If linked with nlopt, this is a GTest which constains several tests:

  • nlopt_serial – A native two parameter nlopt implementation
  • nlopt_op – The same test implemented using the op::NLopt interface
  • nlopt_op_plugin – The same implementation as nlopt_op, but using the dynamic plugin interface.
  • nlopt_op_mpi – A two-rank mpi implementation of the same problem using op‘s “advanced” registration procedure.
  • nlopt_op_mpi_1 – A single-rank implementation using op‘s “advanced” communication pattern. The purpose of this example it to make sure the “advanced” pattern can be used as part of migration to the parallel simulation setting.
  • nlopt_op_bridge – A “black-box” optimizer example using an externally loaded plugin. The external plugin is a custom implementation on ipopt.

License

see LICENSE and NOTICE

SPDX-License-Identifier: BSD-3-Clause

LLNL-CODE-831995

GitHub

View Github