Presentation is loading. Please wait.

Presentation is loading. Please wait.

A survey of Exascale Linear Algebra Libraries for Data Assimilation

Similar presentations


Presentation on theme: "A survey of Exascale Linear Algebra Libraries for Data Assimilation"— Presentation transcript:

1 A survey of Exascale Linear Algebra Libraries for Data Assimilation
Giuliano Laccetti, Marco Lapegna University of Naples Federico II ABSTRACT: Over 70% of real world applications are based on Linear Algebra operations on vectors and matrices residing in core modules and performing time consuming computations. For a such reason several Linear Algebra Libraries are developed since the ’70s and still new and large projects are targeting this scientific area on new HPC systems. The advantages were manifold: the availability of high quality Linear Algebra librarieson exascale systems and software environments allow computational scientists to focus the attention on their own application, without taking into account issues related to the underlying computing environments. On the other hand, since the above mentioned applications mainly spent their time doing linear algebra operations, application developers could obtain maximal performance and accuracy by using this standard library. Additionally, the widespread use of Linear Algebra libraries made it easy to exchange/port high-level libraries, again offering optimal performance. Data Assimilation problems is a typical case in this sense: it is based on several Linear Algebra kernels such as Lanczos iterative algorithms, preconditioning, covariance computation and related factorizations, so that they can take advantage from the availability of recent and ongoing projects like MAGMA, NLAFET and high level computing environments like PETSc. The high modularity of these software tools make possibile the solution of Data Assimilation problems on very different platforms as well as hybrid architectures. Data Assimilation Problems Newton solvers for Non Linear Problems (Jacobian/Adjoint Free Newton Krylov solvers) Lanczos iterative algorithms Preconditioning (spectral / Ritz pairs) covariance computation ScaLAPACK PBLAS CUDA / Xeon PHI BLACS/ MPI Pthreads / OpenMP LAPACK BLAS for specific platform (Intel MKL, AMD ACML) parallel distributed shared sequential low level tools CUDA NOT YET DEFINED ScaLAPACK PETSc NLAFET MAGMA new linear algebra software targeted on new heterogeneous/hybrid "Multicore+GPU" architectures. Supported platforms Intel and AMD multicore CPUs NVIDIA GPU and Intel Xeon Phi accelerators Driver routines main linear algebra factorization Least Squares problems Eigenvalues and Singular Value Decomposition Interfaces for LAPACK and BLAS classical linear algebra software, still a reference point for distributed memory architectures. Based on BLAS/ PBLAS/LAPACK for the computation BLACS/MPI for the communications Driver routines main linear algebra factorization Least Squares problems Eigenvalues and Singular Value Decomposition MAGMA can be integrated to target GPU and multicore devices (to be verify) high level, well known environment for data structures and routine for parallel solution of applications based on PDE Based on BLAS and LAPACK for the computation MPI for the communications Driver routines main linear algebra factorization Newton based nonlinear systems preconditioners sparse and dense matrices, several data structure formats ongoing UE-funded project for linear algebra software on exascale HPC heterogeneous systems Based on To be defined Driver routines dense linear system and eigenvalues direct solution of sparselinear systems communication optimal algorithms for iterative methods


Download ppt "A survey of Exascale Linear Algebra Libraries for Data Assimilation"

Similar presentations


Ads by Google