BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//adass2023//talk//8SB3FP
BEGIN:VTIMEZONE
TZID:MST
BEGIN:STANDARD
DTSTART:20000101T000000
RRULE:FREQ=YEARLY;BYMONTH=1
TZNAME:MST
TZOFFSETFROM:-0700
TZOFFSETTO:-0700
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-adass2023-8SB3FP@pretalx.com
DTSTART;TZID=MST:20231108T084500
DTEND;TZID=MST:20231108T090000
DESCRIPTION:We ported to the GPU with CUDA the Astrometric Verification Uni
 t–Global Sphere Reconstruction (AVU–GSR) Parallel Solver\, developed f
 or the ESA Gaia mission\, by optimizing a previous OpenACC porting of the 
 code. The code finds\, with a 10-100 μas precision\, the astrometric para
 meters of ∼10^8 sources\, the attitude and instrument settings of the Ga
 ia satellite\, and the parameter γ of the PPN formalism\, by solving a sy
 stem of linear equations\, A×x=b\, with the LSQR iterative algorithm. The
  coefficient matrix A of the final Gaia dataset is large\, with ∼10^11x(
 5x10^8) elements\, and sparse\, reaching a size of ∼10–100 TB\, typica
 l for the Big Data analysis\, which requires an efficient parallelization 
 to obtain scientific results in reasonable timescales. In the matrix size\
 , 10^11 is the number of equations\, i.e.\, of stellar observations\, and 
 5x10^8 is the number of unknowns\, Nunk. The speedup of the CUDA code over
  the original AVU–GSR solver\, parallelized on the CPU with MPI+OpenMP\,
  increases with the system size and the number of resources\, reaching a m
 aximum of 14x\, >9x over the OpenACC code. This result is obtained by comp
 aring the two codes on the CINECA cluster Marconi100\, with 4 16 GB V100 G
 PUs per node. We verified the agreement between CUDA and OpenMP solutions 
 for a set of production systems. The CUDA code was then put in production 
 on Marconi100\, essential for an optimal AVU–GSR pipeline and the succes
 sive Gaia Data Releases. We aim to port the production of this code on Leo
 nardo CINECA infrastructure\, expecting to obtain even higher performances
 \, since this platform has 4x GPU memory per node compared to Marconi100.\
 nTo solve a system of linear equations\, the system solution\, the errors 
 on the unknowns (variances) and the covariances can be calculated. Whereas
  the solution and the variances arrays have size Nunk~5x10^8\, the varianc
 es-covariances matrix has a size ~Nunk^2/2\, which can occupy ~1 EB. This 
 represents a “Big Data” problem\, which cannot be solved with standard
  methods. To cope with this difficulty\, we define a novel I/O- based stra
 tegy in a two jobs-pipeline\, where one job is dedicated to the files writ
 ing and the second concurrent job reads the files as they are created\, it
 eratively computes the covariances\, and deletes the files\, to avoid stor
 age issues. In this way\, the covariances calculation does not significant
 ly slowdown the AVU-GSR code for a number of covariances up to ~10^6.\nThe
 se analyses represent a first step to understand the (pre-)Exascale behavi
 or of a class of codes based on the same structure of this one.\nAcknowled
 gments: This work is supported by the Spoke 1 “FutureHPC & BigData” of
  the ICSC– CN di Ricerca in HPC\, Big Data and Quantum Computing–and h
 osting entity\, funded by European Union–NextGenerationEU”. This work 
 was also supported by ASI grant No. 2018-24-HH.0\, in support of the Itali
 an participation to the Gaia mission\, and by CINI\, under the project EUP
 EX\, EC H2020 RIA\, EuroHPC-02-2020 grant No. 101033975.
DTSTAMP:20260712T114156Z
LOCATION:Talks
SUMMARY:The Gaia AVU–GSR Parallel Solver: CUDA solutions for linear syste
 ms solving and covariances calculation toward Exascale infrastructures - V
 alentina Cesare
URL:https://pretalx.com/adass2023/talk/8SB3FP/
END:VEVENT
END:VCALENDAR