Solving Systems of Linear Equations by GPU-based Matrix Factorization in a Science Ground Segment - Archive ouverte HAL Access content directly
Journal Articles Astronomy and Computing Year : 2013

Solving Systems of Linear Equations by GPU-based Matrix Factorization in a Science Ground Segment

(1) , (2) , (3) , (2)
1
2
3

Abstract

Recently, Graphics Cards have been used to offload scientific computations from traditional CPUs for greater efficiency. This paper investigates the adaptation of a real-world linear system solver, which plays a central role in the data processing of the Science Ground Segment of ESA's astrometric Gaia mission. The paper quantifies the resource tradeoffs between traditional CPU implementations and modern CUDA based GPU implementations. It also analyses the impact on the pipeline architecture and system development. The investigation starts from both a selected baseline algorithm with a reference implementation and a traditional linear system solver and then explores various modifications to control flow and data layout to achieve higher resource efficiency. It turns out that with the current state of the art, the modifications impact non-technical system attributes. For example, the control flow of the original modified Cholesky transform is modified so that locality of the code and verifiability deteriorate. The maintainability of the system is affected as well. On the system level, users will have to deal with more complex configuration control and testing procedures.
Not file

Dates and versions

hal-00876195 , version 1 (24-10-2013)

Identifiers

Cite

Maxime Legendre, Albrecht Schmidt, Saïd Moussaoui, Uwe Lammers. Solving Systems of Linear Equations by GPU-based Matrix Factorization in a Science Ground Segment. Astronomy and Computing, 2013, 3-4, pp.58-64. ⟨10.1016/j.ascom.2013.11.004⟩. ⟨hal-00876195⟩
171 View
0 Download

Altmetric

Share

Gmail Facebook Twitter LinkedIn More