High Performance Finite Volume Methods solver for multi-scale cell simulations
Multiscale cell simulators are among the most computationally demanding bioinformatics applications and a source of daunting challenges for supercomputing. Human digital twins require extremely massive simulations that the state-of-the-art is not able to model. Current state-of-the-art systems just...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/403850 |
| Acceso en línea: | https://hdl.handle.net/2117/403850 |
| Access Level: | acceso abierto |
| Palabra clave: | Finite Volume Method High performance computing Bioinformatics Digital Twins High Performance Computing Multi-scale Cell simulation Tridiagonal equations systems Bioinformatics applications Supercomputing Càlcul intensiu (Informàtica) Bioinformàtica Àrees temàtiques de la UPC::Informàtica::Aplicacions de la informàtica::Bioinformàtica |
| Sumario: | Multiscale cell simulators are among the most computationally demanding bioinformatics applications and a source of daunting challenges for supercomputing. Human digital twins require extremely massive simulations that the state-of-the-art is not able to model. Current state-of-the-art systems just reach to real-sized tissues simulations. Physics-based multiscale cell simulator, PhysiCell, sets the basis of a community-built software that helps researchers bridge the intracellular mechanisms to tissue-level biomedical solutions. PhysiCell uses a Finite Volume Method to model the diffusion equation. If we aim at simulating real-sized tissues and as these tools are memory-bound, we need to incorporate MPI into the software. By doing so, the diffusion-decay solver cannot solve the Tridiagonal matrix algorithm in a scalable manner. This work presents a scalable solution, called BioFVM-B, for the diffusion-decay solver in three di- mensions that is decomposed through a Locally One-Dimensional method that accelerates the solution by a first-order splitting in the x-, y- and z-directions. BioFVM-B is presented as a scalable dis- tributed library to model microenvironment evolution with optimized methods for High Performance platforms. It is a performance upgrade of the cutting-edge BioFVM's distributed version, BioFVM- X. The solution included in BioFVM-B involves lightweight microenvironment data structures that enhance memory usage and a new computation workflow to solve a massive number of large tridiago- nal equations systems. Tridiagonal matrix algorithm resolution serialization is concealed by enabling concurrent communication and computation of the different subsets of large tridiagonal systems. Fur- thermore, Auto-fitter is proposed as a pipeline to assess cluster-specific optimal number of subsets that the optimization processes in parallel. The tool uses empirical data from performance tests to select an energy efficient number of steps and nodes for a determined problem size. BioFVM-B allows simulating microenvironments able to contain a real-size tumours by efficiently using up to 4608 cores as a further step towards reaching the virtual digital twins goal. |
|---|