u2.08.06 Parallelism user manual#

Summary:

Any Code_Aster simulation can benefit from the performance gains that parallelism provides (HPC [1] _ in English). These gains can be of two types: on the calculation time and on the space RAM /disk required (per core allocated). Well understood and used well, the impact of HPC on studies can be major in terms of:

Feasibility: unblock a critical situation (model too fine, simulation too slow…);
Productivity: its accelerations up to X10 or even X100 [2] _

contribute to streamlining and securing the conduct of simulations;

Predictivity: allow finer models, more complex and better validated simulations (calibration, sensitivity, parametric…);
Prospective: allow yourself completely different studies.

Code_Aster offers various parallel strategies to adapt to the study and the calculation platform. Some are more focused on computer aspects (distribution of complete calculations or independent modal calculations/ MISS3D, construction of linear systems, basic linear algebra operations), others are more algorithmic (linear solvers HPC MUMPS and PETSc), others are more algorithmic (linear solvers and).

This document briefly describes the organization of parallelism in code. He then recalls some fundamentals in order to help the user take advantage of these parallel strategies. We then detail their implementations, their scope of use and their potential gains. The chains/combinations of different strategies (often natural and set by default) are also discussed.

Users in a hurry can immediately refer to Chapter 2 (« Parallelism in a few clicks! »). It summarizes the procedure for implementing the parallel strategy recommended by default.

Note:

To use Code_Aster in parallel, (at least) three scenarios can occur:

We have access to the centralized Aster machine and we want to use the Astk access interface,
We perform calculations on a cluster or on a multi-core machine with Astk,
Same as the previous case but without Astk.

. toctree:

hidden:

Self