#1333 ISTL GMRES parallel performance bug
Metadata
Property | Value |
---|---|
Reported by | Tobias Malkmus (tomalk@mathematik.uni-freiburg.de) |
Reported at | Aug 20, 2013 19:31 |
Type | Bug Report |
Version | 2.2 |
Operating System | Unspecified / All |
Description
In parallel, the ISTL GMRES implementation calls in each iteration step_sp.dot( ... ) for each Krylov-basisvector, which is at most "number of Restarts"-times.
Usually, a global summation over each process is called in each _sp.dot( ... ) evaluation.
This makes at most "number of Restarts" + 3 numbers of global communication per iteration. In my personal code the parallel performance breaks down for more than 2 processes.
At the moment i don't have an example for the performance break down. I will try to provide one tomorrow.
Best Tobias