Zoltan_LB_Partition erroring on incorrect NUM_GLOBAL_PARTS

[benfulton]

I am getting an error using Zoltan with four nodes, 64 cores per node. I call

Zoltan_set_Param(zz, "NUM_GLOBAL_PARTS", int2str(target_procs))

to make sure that each instance sets the parameter to 256 (target_procs is a command line parameter to the application), but when Zoltan_LB_Partition is called it tells me

Sum of NUM_LOCAL_PARTS 256 < NUM_GLOBAL_PARTS 512

I can't find anywhere where the 512 might have come from. Is it something Zoltan might set on the fly, or perhaps my code is accidentally overwriting a Zoltan variable? I am using Zoltan, not Zoltan2, from Trilinos 16.1.0.

[cgcgcg]

@egboman Could you have a look?

[egboman]

This is odd. In Zoltan, you can set either NUM_LOCAL_PARTS or NUM_GLOBAL_PARTS but not both. The error message indicates both were set, and inconsistently. How many MPI ranks do you have? One per node or one per core? What's your communicator?

[benfulton]

One rank per core, and I am setting both - NUM_LOCAL_PARTS to 1 so each of the 256 ranks assigns one part. Should I not be setting one or the other? The communicator is MPI_COMM_WORLD and I am using OpenMPI.

[egboman]

It is not recommended to set both, so you may have found a bug... Try just set one parameter or the other, not both.