Page 1 of 1

Solving large problems with FETI

Posted: 28 Oct 2011, 10:55
by raback
Hi

There is a rather new FETI (Finite element tear and interconnect) domain decomposition method in Elmer that may help in the solution of large of certain large problems. Below are some scalability results (by Juha) that demonstrate rather reasonable scalability.

The problem under study was an academic case of linear elasticity. Weak scaling was studied so that the number of elements for each partition remained at constant 8000. As the largest case was run with 3375 cores this translates to a problem size of abour 80 million unknowns.

Code: Select all

#procs time(s)  #iter
==================
27    10.52   26
64    12.30   29
125    9.27   31
216    9.96   31
343   10.26   32
512   11.18   32
729   12.13   33
1000  19.88   33
3375  31.52   35
The current implementation involves the solution of a coarse problem of size #procs with one core. This is probably the most important factor behind the increase in simulation time (3375/8000 is not insignificant).

FETI method is very robust and may also be used for the Poisson equation. Implementation of FETI for other equations (Helmholtz or Stokes) is possible but requires additional work.

-Peter

PS. The work was supported by PRACE project and the simulations were carried out on Curie supercomputer at CEA.