severly slow partitioning with v8.6 compared to v8.5

[JoGe]

Dear FEAP community,

I am dealing with a ~ 560,000 nodes problem with 3 DOFs each.
The (serial) partitioning takes 86 sec with FEAPv8.5 but 1709 sec with FEAPv8.6.

Has anyone else faced this issue or is that due to my installation?

Best regards
JoGe

[Prof. S. Govindjee]

Can you put some timing code into both versions to determine which part of the partitioning routines are causing the slowdown?
both graph and outd are in pmacr7.F in the parfeap folder.

[JoGe]

Thank you, Prof. Govindjee for your response!

Please see the attached file for the results of the this problem: an input file with 864000 Hex8 elements was serial partitioned (Metis) using v8.5 and v8.6 for 700 CPUs.

[Prof. S. Govindjee]

What I was hoping for was some function level timing.   Notwithstanding,
can you send me the Input file or a link to the Input file, if it is very large and the
exact commands you are trying to use to partition the Input file.

[arktik]

I am attaching the input file (Iinput) which was used for partitioning and comparison posted above. Please download it (5 days valid) from http://fex.dlr.de/fop/KleMSvXL/test.zip

We can surely do the function level timing as well. Please let us know once the issue is not apparent with some preliminary checks. (The input file simulates a three-point bending beam test in 3D. Material model is replaced with isotropic linear elastic for testing).

[Prof. S. Govindjee]

Thanks.  I'll have a look.

[Prof. S. Govindjee]

Quick update:
1. I have reproduced the issue.
2. Interestingly the graph partitioning is faster in 8.6
3. Even the time from start coord to end elemt is faster in 8.6

4. BUT the time between moving from one file to the next is ~5 seconds in 8.6; in 8.5 it is ~0.1 seconds!

Something strange is going on; I will look deeper.

[Prof. S. Govindjee]

Further update:

In version 8.6 we enhanced the code to allow for linking of dofs and this complicated the determination of which equations belong to which partition (since we partition by node and not equation).  The routine ueqnb1( ) in version 8.6 is the cause of the slowdown that you are seeing.  In version 8.5, on your problem, this set only takes 0.01 seconds on my computer but it requires 4.0+ seconds with version 8.6 --- so a slowdown of 400x.

With the problem identified, we'll see if the algorithm can be sped up.

[Prof. S. Govindjee]

I've pushed a fix for this problem to the download site: version 8.6.1i

This minor release has a few other bug fixes; if you only want this patch, you can just copy parfeap/pmacr7.F.
The patch involves adding back the old method for determining the number of on and off processor equations
when there are no links in the problem.

Let us know how it works for you; it has only been lightly tested

[Prof. S. Govindjee]

Two quick additional comments:

1.  You have a tie command in your file but there is nothing to tie in your file; note that partitioning should only be done on flat files.
2.  Your outd command asks for blocking of block hint size 4 which seems odd for a 3 dof per node problem.  In my testing I just used 'outd aij 1', which grabs its blocking hint size from the number of dofs per node (3 in your case).

[arktik]

Thanks Prof. Govindjee for your help. You are right. TIE command is not doing anything in the input file. The input file is already a flat file created with outmesh. About DOFs, the original problem had 4 dofs. I changed the material model before uploading the files for your testing, but forgot to take care of outd command.

[arktik]

I reran the same problem after updating parfeap/pmacr7.F in v8.6. The problem of slow partitioning is resolved. In fact, v8.6 was ~5-6s faster than v8.5 (for the tested problem). Thanks a lot for your support!