Jul 29 – 31, 2024 In-Person Event
Jülich Supercomputing Centre
Europe/Berlin timezone

Exascale Workflow for NekRS

Not scheduled
20m
Jülich Supercomputing Centre

Jülich Supercomputing Centre

Short Talk

Speaker

Paul Fischer

Description

We discuss two related tools for exascale workflows with NekRS.  The first
is parRSB, which is the parallel partitioner for Nek5000/RS that is based
on recursive coordinate and recursive spectral bisection.  The second is a
new, communication-minimal, coarse-grid solve that is targeting exascale
platforms such as Frontier/Aurora.  The pressure solvers in Nek5000/RS rely
on p-multigrid (pMG) with multilevel Schwarz smoothers for preconditioning
an outer GMRES iteration.  The coarsest level in pMG corresponds to
the unstructured mesh comprising p=1 trilinear elements.  For such meshes,
the number of unknowns is approximately equal to the number of elements, E, and
the number of unknowns per MPI rank is therefore approximately E/P.   On
GPU-based platforms, E/P is typically around 10,000, which is too small to fully
occupy a GPU.  Nek5000/RS mesh sizes, however, can be as large as E = 1 billion,
so the over-all coarse problem is not small.  To reduce communication in the
coarse solve, we employ a two-level Schwarz method with a local solve that
requires only a single pair of nearest-neighbor communications and with a
global reduced-system solve that requires only 2 log_2 P messages. On
Frontier, the communication for the reduced-system solve is on par with a single MPI-allreduce.

Relevance for Nek [100 words max]

Paul

Primary author

Co-authors

Presentation materials

There are no materials yet.