RT info:eu-repo/semantics/article
T1 EPSILOD: efficient parallel skeleton for generic iterative stencil computations in distributed GPUs
A1 Castro, Manuel de
A1 Santamaria Valenzuela, Inmaculada
A1 Torres de la Sierra, Yuri
A1 González Escribano, Arturo
A1 Llanos Ferraris, Diego Rafael
K1 Distributed memory
K1 GPU
K1 Heterogeneous
K1 Stencil
K1 Parallel skeletons
K1 12 Matemáticas
K1 1203.17 Informática
AB Iterative stencil computations are widely used in numerical simulations. Theypresent a high degree of parallelism, high locality and mostly-coalesced memoryaccess patterns. Therefore, GPUs are good candidates to speed up their computa-tion. However, the development of stencil programs that can work with huge grids indistributed systems with multiple GPUs is not straightforward, since it requires solv-ing problems related to the partition of the grid across nodes and devices, and thesynchronization and data movement across remote GPUs. In this work, we presentEPSILOD, a high-productivity parallel programming skeleton for iterative stencilcomputations on distributed multi-GPUs, of the same or different vendors that sup-ports any type of n-dimensional geometric stencils of any order. It uses an abstractspecification of the stencil pattern (neighbors and weights) to internally derive thedata partition, synchronizations and communications. Computation is split to betteroverlap with communications. This paper describes the underlying architecture ofEPSILOD, its main components, and presents an experimental evaluation to showthe benefits of our approach, including a comparison with another state-of-the-artsolution. The experimental results show that EPSILOD is faster and shows goodstrong and weak scalability for platforms with both homogeneous and heterogene-ous types of GPU
PB Springer
SN 0920-8542
YR 2023
FD 2023
LK https://uvadoc.uva.es/handle/10324/58515
UL https://uvadoc.uva.es/handle/10324/58515
LA eng
NO The Journal of Supercomputing, 2023.
NO Producción Científica
DS UVaDOC
RD 21-may-2024