The core library of the Tenzing project. tenzing-core provides facilities for interacting with CUDA + MPI programs as sequential decision problems. This facilitates optimizing CUDA + MPI programs using sequential decision strategies.

Two solvers are available

• tenzing-mcts: Uses Monte-Carlo tree search
• tenzing-dfs: Uses depth-first search

On a supported platform:

source load-env.sh

In any case:

mkdir build && cd build
cmake .. -DCMAKE_CUDA_ARCHITECTURES=70
make

Tests are split into two locations:

• unit tests may be defined in source files
• tests with a more "itegration" flavor are in test/

To run tests, you can do

• make test
• ctest
• tenzing-all
  • -ltc: list tests cases
  • -tc="a,b": only run test cases named a and b

This creates some CMake complexity, as the test functions present in static libraries will not be linked into the resulting test binary. Therefore, we use a CMake object library to generate the test binary, and then generate a static library from the object library. object library properties do not get propagated properly / at all, so we have to redefine what needs to be linked and included, etc

tenzing-core has been tested on the following platforms:

• NERSC perlmutter: g++ 10.3 / nvcc 11.4 / Cray MPICH 8.1.13
• Sandia vortex (similar to ORNL Lassen and OLCF Summit): g++ 7.5.0 / nvcc 10.1 / IBM Spectrum MPI
• Sandia ascicgpu

• Visit the API documentation in docs/api.md
• ascicgpu system documentation in docs/ascicgpu.md
• vortex system documentation in docs/vortex.md
• perlmutter ssytem documentation in docs/perlmutter.md

• python bindings (with pybind11)

See CONTRIBUTING.md for contribution guidelines.

• enable / disable CUDA / MPI
  • isolate Ser/Des
  • isolate platform assignments
• a BoundOp cannot produce the std::shared_ptr<OpBase> of it's unbound self, only OpBase
  • can't ask an std::shared_ptr<BoundOp> for std::shared_ptr<OpBase>
  • maybe std::shared_from_this?
• special status of Start and End is a bit clumsy.
  • maybe there should be a StartEnd : BoundOp that they both are instead of separate classes
    • in the algs they're probably treated the same (always synced, etc)
• Platform is a clumsy abstraction, since it also tracks resources that are only valid for a single order
  • e.g., each order requires a certain number of events, which can be resued for the next order

Please see NOTICE.md for copyright and license information.