Multi-GPU implementation of the Levy and Lindenbaum's method for streaming SVD with application to large datasets of KHRTI simulation
Abstract
Singular value decomposition (SVD) is a fundamental technique in data-driven approaches, particularly in fluid mechanics. It enables the decomposition of complex datasets, the extraction of dominant flow structures, and the construction of reduced-order models, thereby significantly reducing the complexity of physical systems. However, when applied to extremely large datasets involving millions of degrees of freedom, traditional SVD methods, even in parallel settings, become computationally prohibitive or infeasible. To overcome this limitation, the innovative approach proposed by Levy and Lindenbaum for computing lowrank approximations can be employed. This technique allows for streaming data processing, significantly reducing computational complexity in both space and time, and enabling the treatment of datasets that far exceed available memory. We propose a new parallel implementation of this method, leveraging multiple GPUs. As a case study, we apply our implementation to data obtained by direct numerical simulation (DNS) of the Kelvin-Helmholtz-Rayleigh-Taylor instability (KHRTI), which comprises 480 million spatial degrees of freedom and 875 snapshots, demonstrating the scalability, robustness, and practical efficiency of our implementation in handling very large datasets.
Related
- Gate-Efficient Implementation of the Query-Optimal Time-Dependent Hamiltonian Simulation
- Generative marginal learning for finite-size molecular simulation with application to stress calculation
- Design, implementation and validation of a web-based dynamic simulation platform for pumped storage hydropower plant control systems
- Multigroup Radiation Diffusion with Compton Scattering on a Moving Mesh: Implementation in RICH and Application to Tidal Disruption Events
- A Differentiable, GPU-Accelerated Finite-Element Formulation in JAX for Large-Deformation Shells with Contact: Application to Patient-Specific Aortic Valve Simulation
BibTeX
@article{ajanif2026multigpu,
title = {{Multi-GPU implementation of the Levy and Lindenbaum's method for streaming SVD with application to large datasets of KHRTI simulation}},
author = {Yassin Ajanif and Lucas Lestandi and Luisa Silva and Aditi Sengupta and Bhavna Joshi and Hugues Digonnet},
journal = {Computer Physics Communications},
year = {2026},
month = aug,
doi = {10.1016/j.cpc.2026.110364},
url = {https://hal.science/hal-05315804v2},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Multi-GPU implementation of the Levy and Lindenbaum's method for streaming SVD with application to large datasets of KHRTI simulation
AU - Yassin Ajanif
AU - Lucas Lestandi
AU - Luisa Silva
AU - Aditi Sengupta
AU - Bhavna Joshi
AU - Hugues Digonnet
PY - 2026
DA - 2026/08/20
JO - Computer Physics Communications
DO - 10.1016/j.cpc.2026.110364
UR - https://hal.science/hal-05315804v2
AB - Singular value decomposition (SVD) is a fundamental technique in data-driven approaches, particularly in fluid mechanics. It enables the decomposition of complex datasets, the extraction of dominant flow structures, and the construction of reduced-order models, thereby significantly reducing the complexity of physical systems. However, when applied to extremely large datasets involving millions of degrees of freedom, traditional SVD methods, even in parallel settings, become computationally prohibitive or infeasible. To overcome this limitation, the innovative approach proposed by Levy and Lindenbaum for computing lowrank approximations can be employed. This technique allows for streaming data processing, significantly reducing computational complexity in both space and time, and enabling the treatment of datasets that far exceed available memory. We propose a new parallel implementation of this method, leveraging multiple GPUs. As a case study, we apply our implementation to data obtained by direct numerical simulation (DNS) of the Kelvin-Helmholtz-Rayleigh-Taylor instability (KHRTI), which comprises 480 million spatial degrees of freedom and 875 snapshots, demonstrating the scalability, robustness, and practical efficiency of our implementation in handling very large datasets.
SN - 0010-4655
SN - 1879-2944
ER -