| |
 |
| Paper: |
The Gaia AVU - GSR Parallel Solver: CUDA Solutions for Linear Systems Solving and Covariances Calculation Toward Exascale Infrastructures |
| Volume: |
541, ADASS XXXIII |
| Page: |
241 |
| Authors: |
Valentina Cesare; Ugo Becciani; Alberto Vecchiato; Mario Gilberto Lattanzi; Fabio Pitari; Giuseppe Tudisco; Marco Aldinucci; Beatrice Bucciarelli |
| DOI: |
10.26624/JDOS5920 |
| Abstract: |
We ported to the GPU with CUDA the Gaia Astrometric Verification
Unit–Global Sphere Reconstruction (AVU - GSR) Parallel Solver by optimizing a previous OpenACC porting of the code. The code finds the astrometric, attitude, calibration, and global parameters of ∼108 sources, by solving a system of linear equations,
with the LSQR iterative algorithm. The coefficient matrix is large, with ∼1011×(5×108)
elements, and sparse, reaching a size of ∼10–100 TB. The speedup of the CUDA code
over the original, CPU-parallelized, AVU - GSR solver reaches ∼14×. A less trivial
problem is provided by the computation of the covariances of the system, which can
occupy∼1 EB. To compute a subset of the covariances, we define a novel I/O-based
pipeline with two jobs, concurrently launched. We aim to port the production of this
code on Leonardo CINECA infrastructure, in perspective of pre-Exascale systems. |
|
|
|
|
|
