betaFeedback

WorksPhysics and Astronomy

Communication-Efficient Hamiltonian Evaluation for Massively Parallel Tensor-Product Bitstring Selected Configuration Interaction

Abstract

Extreme-scale selected configuration interaction (SCI) requires coordinated Hamiltonian evaluation, distributed CI-vector storage, and interprocess communication on massively parallel computers. Our previous work [J. Chem. Phys. 164, 194103 (2026)] introduced the tensor-product bitstring selected configuration interaction (TBSCI) framework, combining on-the-fly Hamiltonian evaluation with fully distributed CI-vector storage and enabling parallel diagonalization of trillion-determinant spaces on up to 54,000 Fugaku nodes. Here, we introduce FLIP, a new Hamiltonian evaluation algorithm within TBSCI that reorganizes the CI-vector layout and excitation-link data to replace remote CI-coefficient transfers for αα-electron double excitations with local coefficient access and link retrieval from nearby MPI ranks. In communication-intensive full configuration interaction (FCI) benchmarks on Fugaku, no exposed communication waiting time was recorded on the slowest process across all 32 measured Hamiltonian-vector multiplications. For N2_2/cc-pVTZ with 2.62×10122.62\times10^{12} determinants, the multiplication achieved 91.3% strong-scaling efficiency from 18,000 to 81,000 nodes (more than 3.8 million compute cores). We also obtained an FCI reference potential-energy curve for N2_2/cc-pVTZ. At all six bond lengths, spin bitstrings selected by their collective weights in reference SCI wavefunctions yielded variational TBSCI energies within 0.1 mEhE_h of FCI using fewer than 1% of the FCI determinants, demonstrating the compactness of the tensor-product bitstring representation. Finally, we discuss local construction of double-excitation links to substantially reduce communication volume and eliminate persistent storage of precomputed double-link tables for future second-order perturbation theory calculations with large external determinant spaces.

Paper details

vlnc:W_pny8hpfy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

ChatGPT (OpenAI)

View statement
Looking for related papers…