metis 5.0 on BluegeneP


I am trying to use metis 5.0 metis_partmeshdual function on a BluegeneP node.
The installation did not worked. I had to change the definition of the variable GKLib in CMakeList.txt:

I also had to add "-DCMAKE_CXX_COMPILER=xlc++_r" to CONFIG_FLAGS in the makefile to define the c++ compiler.
With these changes metis compiled on the bluegeneP and the POWER7.

The code (in Fortran) is working on the power7 and my machine (Linux PC) but not on the bluegeneP.

Here is the call to the function:
call METIS_PartMeshDual
& (nelem2, npoin2, eptr, eind, nulltable,
& nulltable, ncommonnodes, nparts, nulltable,
& nulltable, edgecut, epart, npart)
nulltable is an non allocated integer table.
I checked in metis code and it considered as NULL in C.
I checked the input tables which are ok.
I debugged with printf and ended in the function SetupCtrl of METIS_PartGraphKway.

The table are renumbered to fit a C numbering because using the option table
was not working and it is unfit for Fortran to define this table as we don't have access to the C macro to get the value of each parameter.

Do you have any ideas what is going wrong ?

I am at your disposal for any informations.

Yoann Audouin

RE: metis 5.0 on BluegeneP


I tested the latest release 5.0.1.
And I got the same error.


Yoann Audouin

RE: Did you try to do a 'make

Did you try to do a 'make config CC=' to specify the c compiler to use?
Did the cmake build worked on your linux PC?

What is the thing that is going wrong?

RE: Hi, I used CC= to define the


I used CC= to define the C compiler but cmake was still choosing /usr/bin/g++ for the C++ compiler.
The cmake worked on my linux PC.

Are there any tests in metis source that I could run to see if it is working on the bluegeneP ?

About the execution I have two problem:
- Do you have another way to pass null pointer for the parmater than the nulltable I am using
because the nulltable solution is not portable and could easily leads to bug.
- The metis function just dump a core I am working on locating the error I will give you more information as soon as I have them.



RE: Re: metis 5.0 on BluegeneP


I run he code on gdb and I got this :
0x011636e8 in libmetis__SetupCtrl (optype=METIS_OP_KMETIS, options=0x0, ncon=1, nparts=8, tpwgts=0x0, ubvec=0x0) at /gpfs/stfc/exccp12a/ezg07126/PARTEL_PARTIONNER/metis-5.0/libmetis/options.c:107
107 ctrl->tpwgts[i*ncon+j] = 1.0/nparts;
This happened for i=0 j=0.

I hope this will help