Draft: INT: Integration of gmresSolver from foam-extend
Summary
Integration of Foam-extend (http://www.foam-extend.org) gmres solver
Details of new models (If applicable)
- should work standalone or in combination with GAMG (with processor agglomeration)
- requires number of orthogonal directions (nDirections) as input
Risks
- will not help with scaling - no pipelining
- uses more memory (scales with nDirections)
Edited by Kutalmış Berçin
Merge request reports
Activity
assigned to @andy
requested review from @kuti
We have observed the following issues in the parallel execution of the GMRES (using
scotch& 2 processors):- Unlike the serial runs, the residuals indicate the divergence of the simulation.
- The simulation takes at least two orders of magnitude more time in comparison to the PCG case.
Run
Allrun-parallelin longPipe-GMRES-04Jun24.zip, and see if the same observations are made.The issues seem to be improving for the scope of testing once we use the
hierarchicaldecomposition method.Edited by Kutalmış Berçin-
See the initial-residual results obtained from the longPipe-05Jun24.zip:
Mode PCG GMRES Serial 
Parallel (Hierarch. - 2 procs in streamwise dir.) Might the following execution-time data be misleading, yet still indicative:
Mode PCG GMRES Serial [sec] 189.6 1113.4 Parallel (Hierarch. - 2 procs in streamwise dir.) [sec] 157.2 555.5 Edited by Kutalmış Berçin Running with parProfiling and GMRES with 10 nDirections:
log.simpleFoam.PCG log.simpleFoam.GMRES.10
- reductions / waits scale with number of directions
- probably not useful in parallel without some communication hiding / pipelining
Please register or sign in to reply