The project aims to address the development of parallel codes for the solution of large- and huge-scale inverse problems in imaging. The starting points are the codes developed and analyzed in previous ISCRA class C projects ("ParJoInv" and "PANOIP"), which in different ways showed that the challenging inversion problems currently arising in a number of applications fields, such as Geophysics, Medicine, Astronomy, Microscopy, etc., can effectively be faced in reasonable time only with the support of parallel codes for HPC architectures. These problems usually belong to the class of simply constrained nonlinear programming (NLPs) problems and require the minimization of (usually heavy) nonlinear functionals. The functionals are generally composed by one (or more) best-fit term and one (or more) regularization terms. The size of the resulting NLPs overcomes very quickly the tens of millions variables. We will consider first-order iterative methods for the solution, particularly the scaled gradient-projection (SGP) approach. This project is aimed to let the research codes developed so far make a step ahead, by both optimizing the code's structure and by introducing hybrid programming, that is by mixing inter-node distributed-memory computations (the MPI-related part) with intra-node shared-memory multithreaded computations (the OpenMP-related part). Here the computing nodes are thought to be multicore CPUs. This is an higher level of parallelization, which in recent years has shown to be very effective in catching the benefits of both the programming paradigms. It will be part of the project's goals to verify whether and how this hybrid approach is compliant and suitable with PETSc and TAO libraries. An additional goal is the integration of GPUs and multicore CPUs: these configurations are increasingly appealing for their reduced costs and good performances. The implementation of the SGP approach within such a mixed environment is surely possible, but has a number of nontrivial issues to face. The project's potential outcomes are impressive: for instance, we expect to face huge-scale 3D Microscopy imaging problems provided by next-generation devices. Also, we plan to compute solutions of 3D Geophysics reconstructions of very large volumes. PROJECT'S BUDGET: 50000 core hours (the maximum allowed). (CINECA's national ISCRA projects are issued as part of the European PRACE Tier0 access initiative)
Large-scale parallel computing for inverse problems in imaging
ZANGHIRATI, Gaetano
2012
Abstract
The project aims to address the development of parallel codes for the solution of large- and huge-scale inverse problems in imaging. The starting points are the codes developed and analyzed in previous ISCRA class C projects ("ParJoInv" and "PANOIP"), which in different ways showed that the challenging inversion problems currently arising in a number of applications fields, such as Geophysics, Medicine, Astronomy, Microscopy, etc., can effectively be faced in reasonable time only with the support of parallel codes for HPC architectures. These problems usually belong to the class of simply constrained nonlinear programming (NLPs) problems and require the minimization of (usually heavy) nonlinear functionals. The functionals are generally composed by one (or more) best-fit term and one (or more) regularization terms. The size of the resulting NLPs overcomes very quickly the tens of millions variables. We will consider first-order iterative methods for the solution, particularly the scaled gradient-projection (SGP) approach. This project is aimed to let the research codes developed so far make a step ahead, by both optimizing the code's structure and by introducing hybrid programming, that is by mixing inter-node distributed-memory computations (the MPI-related part) with intra-node shared-memory multithreaded computations (the OpenMP-related part). Here the computing nodes are thought to be multicore CPUs. This is an higher level of parallelization, which in recent years has shown to be very effective in catching the benefits of both the programming paradigms. It will be part of the project's goals to verify whether and how this hybrid approach is compliant and suitable with PETSc and TAO libraries. An additional goal is the integration of GPUs and multicore CPUs: these configurations are increasingly appealing for their reduced costs and good performances. The implementation of the SGP approach within such a mixed environment is surely possible, but has a number of nontrivial issues to face. The project's potential outcomes are impressive: for instance, we expect to face huge-scale 3D Microscopy imaging problems provided by next-generation devices. Also, we plan to compute solutions of 3D Geophysics reconstructions of very large volumes. PROJECT'S BUDGET: 50000 core hours (the maximum allowed). (CINECA's national ISCRA projects are issued as part of the European PRACE Tier0 access initiative)I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.