Defesa – Dissertação de Moisés Felipe Lehnen

Aluno(a): Moisés Felipe Lehnen
Orientador(a): Philippe Olivier Alexandre Navaux

Título: Efficient MPI One-Sided Communication Strategies for Dependency-Chained Solvers
Linha de Pesquisa: Computação de Alto Desempenho e Sistemas Distribuídos

Data: 29/09/2026
Hora: 09:40
Local: Esta banca ocorrerá de forma presencial na sala sala 215 predio 43412 do Instituto de Informática/UFRGS.

Banca Examinadora:
-Luciano Paschoal Gaspary (UFRGS)
-Tiago Coelho Ferreto (PUCRS)
-Dalvan Jair Griebler (PUCRS)

Presidente da Banca: Philippe Olivier Alexandre Navaux

Resumo: Many iterative stencil solvers rely on dependency-chained halo exchanges, where sub- domain progress depends on timely neighbor updates, making synchronization overhead a key scalability bottleneck. This work investigates the use of MPI passive-target re- mote memory access (RMA) to decouple data movement from synchronization in such workloads. We redesign the halo exchange of a finite-difference wave-propagation solver using RMA and compare it against a nonblocking two-sided implementation across mul- tiple configurations, including varying problem sizes and process counts. We analyze how communication costs are redistributed when using RMA. Our results show that RMA can reduce synchronization overhead within the time-stepping loop, achieving up to 2.0× speedup in this phase. These gains are partially offset by increased costs in synchronization-heavy phases such as window teardown, highlighting a shift rather than an elimination of overheads. These results highlight when RMA improves scalability and provide guidelines for separating synchronization from data movement in stencil-based solvers.

Palavras-Chave: Memory, Communication, Distributed, Dependency-Chained, Scalability, MPI, RMA, Wavefield, Solver, Exchange