The MPI + CUDA Gaia AVU–GSR Parallel Solver Toward Next-generation Exascale Infrastructures

We ported to the GPU with CUDA Astrometric Verification Unit-Global Sphere Reconstruction (AVU-GSR) Parallel Solver developed for ESA Gaia mission, by optimizing a previous OpenACC porting of this application. The code aims find, [10,100]$\mu$as precision, astrometric parameters $\sim$$10^8$ stars, attitude and instrumental settings satellite, global parameter $\gamma$ parametrized Post-Newtonian formalism, solving system linear equations, $A\times x=b$, LSQR iterative algorithm. coefficient matrix $A$ final dataset is large, $\sim$$10^{11} \times 10^8$ elements, sparse, reaching size $\sim$10-100 TB, typical Big Data analysis, which requires an efficient parallelization obtain scientific results in reasonable timescales. speedup over original AVU-GSR solver, parallelized on CPU MPI+OpenMP, increases number resources, maximum $\sim$14x, >9x This result obtained comparing two codes CINECA cluster Marconi100, 4 V100 GPUs per node. After verifying agreement between solutions set systems different sizes computed OpenMP that showed required was put production essential optimal pipeline successive Releases. analysis represents first step understand (pre-)Exascale behavior class applications follow same structure code. In next months, we plan run pre-Exascale platform Leonardo CINECA, next-generation A200 node, toward infrastructure, where expect even higher performances.