Mamute: high-performance computing for geophysical methods
João B. Fernandes, Antônio D. S. Oliveira, Mateus C. A. T. Silva, Felipe H. Santos-da-Silva, Vitor H. M. Rodrigues, Kleiton A. Schneider, Calebe P. Bianchini, João M. de Araujo, Tiago Barros, Ítalo A. S. Assis, Samuel Xavier-de-Souza · 2025
Read Mamute: high-performance computing for geophysical methods by João B. Fernandes, Antônio D. S. Oliveira, Mateus C. A. T. Silva, Felipe H. Santos-da-Silva, Vitor H. M. Rodrigues, Kleiton A. Schneider, Calebe P. Bianchini, João M. de Araujo, Tiago Barros, Ítalo A. S. Assis, Samuel Xavier-de-Souza (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-12350.txt.
Introduction
A large team of co-authors wrote a paper in two thousand twenty-five presenting software for geophysics on high-performance computers. Earth-science methods such as seismic imaging solve wave equations over huge volumes, which is very costly to compute. Code must squeeze large machines well or the work crawls.
Their program, Mamute, offers seismic modeling and full waveform inversion, a method that fits a model of the earth to recorded waves. It adds fault tolerance, automatic scheduling of parallel loops, and workload balancing across machines. The authors demonstrate both methods running in the program. The code is written in C++ and released under the MIT license.
Catalog record
| CREATOR | João B. Fernandes, Antônio D. S. Oliveira, Mateus C. A. T. Silva, Felipe H. Santos-da-Silva, Vitor H. M. Rodrigues, Kleiton A. Schneider, Calebe P. Bianchini, João M. de Araujo, Tiago Barros, Ítalo A. S. Assis, Samuel Xavier-de-Souza |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-12350
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Mamute: high-performance computing for geophysical methods João B. Fernandes, Antônio D. S. Oliveira, Mateus C. A. T. Silva, Felipe H. Santos-da-Silva, Vitor H. M. Rodrigues, Kleiton A. Schneider, Calebe P. Bianchini, João M. de Araujo, Tiago Barros, Ítalo A. S. Assis, Samuel Xavier-de-Souza Abstract Due to their high computational cost, geophysical applications are typically designed to run in large computing systems. Because of that, such applications must implement several high-performance techniques to use the computational resources better. In this paper, we present Mamute, a software that delivers wave equation-based geophysical methods. Mamute implements two geophysical methods: seismic modeling and full waveform inversion (FWI). It also supports high-performance strategies such as fault tolerance, automatic parallel looping scheduling, and distributed systems workload balancing. We demonstrate Mamute’s operation using both seismic modeling and FWI. Mamute is a C++ software readily available under the MIT license. Journal: SoftwareX Nr. Code metadata description Metadata C1 Current code version 1.0 C2 Permanent link to code/repository used for this code version https://github.com/lapps-ufrn/mamute C3 Permanent link to Reproducible Capsule https://github.com/lapps-ufrn/mamute/blob/master/doc/mamute/quick-start.md C4 Legal Code License MIT License. C5 Code versioning system used git C6 Software code languages, tools, and services used C/C++, Python, OpenMP, MPI. C7 Compilation requirements, operating environments & dependencies CMake 3.5, GCC 8 C8 If available Link to developer documentation/manual https://github.com/lapps-ufrn/mamute/tree/master/doc C9 Support email for questions [email protected] Table 1 : Code metadata. 1 Motivation and significance Many applications employed by the geophysics community in the search for oil and gas reservoirs are based in the solution of the wave equation. Full waveform inversion (FWI) is one of the main geophysics methods [ 1 ] . It consists of an inverse problem [ 2 ] , wherein seismic data observations are leveraged to glean insights into the Earth’s subsurface. FWI’s inverse problem is solved with numerical optimization methods, which employ the misfit between the observed seismograms and modeled data to obtain its target information. The modeled data in FWI is mostly generated by solving discrete versions of a given wave equation. In this paper, we introduce Mamute , a C++ software designed to execute wave equation based methods such as three-dimensional seismic modeling and FWI. Methods based on the solution of the wave equation typically require significant time and computational resources. This is primarily due to the need to apply numerical solutions of partial differential equations (PDEs) and optimization methods on extensive datasets. To enhance performance, Mamute employs parallel programming techniques, leveraging a combination of OpenMP and MPI. Additionally, Mamute is equipped with various tools optimized for modern high-performance computing (HPC) setups. These include workload scheduling techniques for both distributed [ 3 ] and shared [ 4 ] memory environments, fault tolerance mechanisms [ 5 ] , and optimal checkpointing methods [ 6 , 7 ] , for efficient data modeling and gradient evaluation. These features collectively ensure Mamute ’s seamless execution on contemporary supercomputers. Many of these HPC capabilities have been the focus of numerous research projects. Fernandes et al. [ 8 ] introduced a distributed task scheduling algorithm tailored for Mamute . Da Silva et al. [ 9 , 10 ] explored an auto-tuning approach to schedule OpenMP tasks in Mamute dynamically. Santana et al. [ 11 ] validated a fault tolerance mechanism within Mamute . 2 Software description Mamute is a C++ program that applies HPC techniques to improve seismic applications. The code is hosted in GitLab and open to contribution with MIT Licence. 2.1 Software functionalities Mamute is a software suite that implements algorithms based on the numerical solution of the acoustic wave equation in the time domain. Mamute solves the 3D acoustic wave equation through the finite differences method with a second-order approximation in time and an eighth-order approximation in space [ 12 ] . Optionally, the wave equation can consider the density influence [ 13 ] . The acoustic wave equation and the variable-density wave equation in a heterogeneous medium are given, respectively, by: \frac{\partial^{2}p}{\partial t^{2}}=c^{2}({\mathbf{x}})\nabla^{2}p+S({\mathbf{x}},t)\quad\mbox{and}\quad\frac{\partial^{2}q}{\partial t^{2}}=c^{2}({\mathbf{x}})(\nabla^{2}q-m^{2}({\mathbf{x}})q)+\frac{S({\mathbf{x}},t)}{\sqrt{\rho({\mathbf{x}})}}, (1) where {\mathbf{x}}\in\Omega\subset\mathbb{R}^{3} and t\in\mathbb{R} are, respectively, the Cartesian spatial coordinates and the time, p({\mathbf{x}},t) is the acoustic pressure, q({\mathbf{x}},t)=p({\mathbf{x}},t)/\sqrt{\rho({\mathbf{x}})} is the normalized wavefield, being \rho({\mathbf{x}}) the material density, c({\mathbf{x}}) is the P-wave velocity of the subsurface, m^{2}({\mathbf{x}})=\nabla^{2}(\sqrt{\rho({\mathbf{x}})})/\sqrt{\rho({\mathbf{x}})} represents the mass term. The source term S({\mathbf{x}},t) is often represented as f(t)\delta({\mathbf{x}}-{\mathbf{x}}_{S}) , where f(t) is the time-dependent source signature, {\mathbf{x}}_{S} denotes the source location, and \delta({\mathbf{x}}-{\mathbf{x}}_{S}) is the Dirac delta function used to localize the source in space. Based on that wave propagation operator, Mamute has implemented two seismic processing algorithms, the seismic modeling [ 14 ] and the full waveform inversion (FWI) [ 1 , 2 ] . For the moment, all float-point operations on Mamute are performed with double precision. Mamute ’s seismic modeling features absorbing boundaries implemented via damping [ 15 , 16 ] and convolutional perfectly matched layer (CPML) [ 17 ] . Both techniques are used to allow the simulation of the wave propagation on Earth using a limited computational mesh. Mamute ’s user can also choose to perform the wave propagation with free surface, i.e. , simulating the wave reflections of the surface. Furthermore, time interpolation via the nearest neighbor algorithm or sinc function can be employed to resample seismic data. Using a text configuration file, the user can define the seismic acquisition geometry, including the sources’ and receivers’ positions, the velocity model’s file and its discretization, the source signature file and its peak frequency and amplitude, the absorbing boundary thickness, the number of time steps and the time resolution. Mamute ’s FWI features L-BFGS-b [ 18 , 19 ] as its optimization method. Gradient preconditioning [ 20 ] is used to smooth the gradient. This software also qualifies for fault tolerance through the Dependability Library for Iterative Applications (DeLIA) [ 21 ] , which offers fault detection, data conservation, and failover capabilities. By doing so, the FWI can restart from the state of the execution at a time before the execution is interrupted. The gradient is computed using the adjoint-state method [ 22 ] . Using a text configuration file, the user can define a limit of memory usage for checkpointing, the L-BFGS-b’s boundaries and maximum number of iterations, the maximum number of velocity model updates, and the type of gradient preconditioning (no filter, Bessel filter, or Laplace filter) and its filter coefficients. The FWI computes the wavefield in forward and reverse time. Forward data needs to be reused when computing the reverse wavefield. In those circumstances, a high amount of storage is needed to store the forward wavefield. Mamute has three options for managing the forward wavefield. The user can compile Mamute to save the entire wavefield in memory, improving the adjoint performance but consuming a high amount of main memory. For large problems, the RAM may not be enough. An alternative is to save the forward wavefield in the disk. This option will decrease main memory usage considerably, but it will require time to save and read the wavefield from secondary storage. The third choice is to use the optimal checkpointing [ 6 ] technique to save selected timestamps from the forward wavefield and then recompute just the missing data. Mamute ’s seismic modeling and FWI are implemented with two levels of parallelization. Inter-node parallelization