Large-Scale Hierarchical Time-Series Reconciliation

hts.ml.works

Building

1. Install Eigen 3 for OpenMP / LAPACK-like support on matrix operations
   • use -DCMAKE_INSTALL_PREFIX= to install locally
   • export CMAKE_PREFIX_PATH="$CMAKE_PREFIX_PATH:{YOUR_PATH_TO_EIGEN3}" to make sure CMake functions properly
2. Clone this repository with pybind11 submodule
3. Install requirements in requirements.txt
   • Remark our own fork of pytest-mpi that is capable of producing the stdout of a single process while suppressing others.
4. python -m pip install large_scale_hts_reconciliation/ to install the pip package with CMake

Reproducing Benchmarking Results

1. Allocate two GCP e2-highcpu-16 instances with 1 vCPU per core, 8 cores and Intel Broadwell x86/64 platform/architecture. One instance is in us-central1-a and the other is in us-east1-b1.
2. Install OpenMPI 4.1.4 and set up hosts file properly (see passwordless SSH)
3. For single-node results, run mpirun -n 8 --bind-to socket python3 -m pytest --with-mpi $HOME/large-scale-hts-reconciliation/large_scale_hts_reconciliation/tests/test_mpi_fast.py --unmute-ranks=0 and mpirun -n 1 --bind-to socket python3 -m pytest tests/test_single_process_fast.py for multi-process benchmarks (test_mpi_fast.py) and single-process benchmarks (test_single_process_fast.py)
4. For multi-node results, run mpirun --hostfile hosts -n 8 --bind-to socket python3 -m pytest --with-mpi $HOME/large-scale-hts-reconciliation/large_scale_hts_reconciliation/tests/test_mpi_fast.py --unmute-ranks=0
5. For OLS/WLS results, run mpirun -n 1 --bind-to socket python3 -m pytest tests/test_single_process_slow.py. They are quite slow.

LHTS Benchmarks

All benchmarks are available on this GCS bucket: gs://lhts