No speedup observed after using renumberMesh
 

Hi,

I have been using renumberMesh to speed up my parallel calculation of pimpleDyMFoam (for 17 million cells sperad across 256 processors, having AMI interface) but didn't observe any increase in speed compared to a similar case without using renumberMesh.

I used the utility after the parallel decomposition and the bandwidth decreased significantly to 2379 from 67180 after using the utility. Can someone let me know if he/she has obtained any increase in speed using this utility?

I am not sure, but I believe you need to renumber before you decompose?

Also, if you have a large number of processor, the impact of renumbering will be much smaller.

Can anyone confirm this - ie that you should run renumberMesh before you decomposePar??

Thanks in advance

-- Can anyone confirm this - ie that you should run renumberMesh before you decomposePar??

Yes.

Please use "renumberMesh" in serial prior to the decomposition step. But I can't recall the main reason why we should avoid "renumberMesh" in parallel (but there was a reason - if I am not wrong, it was about sampling.)

If you still want to run the "renumberMesh" in parallel, please use its "-parallel" flag.

Hi there,

Thanks for your answer but the question was not whether renumberMesh should be executed in parallel or serial, but rather whether renumberMesh should be run on the undecomposed mesh or the decomposed mesh.

Since renumbering algorithms are designed to move the coefficients closer to the diagonal (reduce coefficient matrix bandwidth) then its not immediately clear whether if you did this before a decomposePar, whether the better conditioned coefficient matrix would be persist (and be optimal) after a domain decomposition.

I tried renumberMesh several times over the years and never had a significant effect, no matter the order of execution.

In my most recent test (4.7 mio cells, 64 procs, interFoam, v2012), I observed no speedup if renumbered before decomposition and even a speed loss if renumbered after decomposition. Even though the bandwidth reported decreased strongly in both cases.

Interesting, good to know, thanks David. I found some reasonanble speed ups when I used renumbering in Fluent combined with the coupled PBS, so perhaps the segregated algorithms in OF are the reason.

I had thought a stronger diagonal in the coefficient matrix might improve stability but from memory, I think thats probably more the conditioning of the matrix i.e. the ratio of the largest / smallest eigen values rather than the bandwidth / diagonal ...