ANSYS Distributed Memory Parallel: MPI Version Mismatch Diagnosis

Direct Answer

For ANSYS /CONFIG processor count and memory-per-core to be interpreted correctly on a cluster scheduler: (1) a DMP license must be available and distinct from the SMP license used locally, (2) the processor count must exactly match the scheduler-allocated ranks (e.g., SLURM_NTASKS or PBS nodefile slots), (3) memory-per-core must not exceed the physical memory per core on the allocated nodes, and (4) the solver launch command must consume the same rank list or machine file that the scheduler provides. The MPI implementation (Intel MPI, MPICH, or OpenMPI) must match the ABI the solver binaries were linked against; mixing them causes MPI_Init failures, incorrect rank-to-node mapping, or wrong domain decomposition. Uniform file access requires every compute node to mount the same shared filesystem (NFS, Lustre, GPFS) at identical absolute paths so the working directory, input mesh, and result files resolve without remapping; node-local scratch is not uniform by default and requires explicit staging.

Likely Explanation vs. Confirmed Facts

Likely explanation: Local SMP validation exercises only single-node shared-memory execution with a local license checkout. It does not test DMP license checkout, scheduler slot mapping, MPI launcher arguments, or cross-node path resolution. The cluster job likely requests DMP but launches with an incorrect -np value, an MPI ABI mismatch, or inaccessible input paths.

Confirmed conditions:

• DMP license must be reachable through any firewall from all compute nodes.
• Processor count in /CONFIG must equal the scheduler's allocated task count.
• Memory-per-core is enforced by the solver; the scheduler's memory limit can independently kill the job.
• Solver binaries are linked to one MPI ABI per ANSYS release; Intel MPI, MPICH, and OpenMPI are generally not interchangeable at runtime.
• All nodes must see identical absolute paths for the working directory and input files.

Minimal Verification Steps

1. On a compute node, run ldd <solver_executable> and compare the MPI library paths against $I_MPI_ROOT or mpirun --version to confirm the loaded module matches the solver's linked MPI.
2. Submit a minimal DMP job with -np equal to the allocated slots; inspect solver output for expected rank count, node list, and domain decomposition summary.
3. Run hostname; pwd; ls -l <inputfile> on two allocated nodes to verify identical absolute path and inode for shared files.
4. Inside the job, check env | grep -i -E 'SLURM|PBS|MPI|I_MPI' to confirm rank count and MPI root match the ANSYS launch command.

One Missing Diagnostic Detail

Provide the exact ANSYS product (Mechanical or Fluent), release version (e.g., 2024 R2), the MPI module and version loaded on the cluster (module list output), and the scheduler type (SLURM, PBS, LSF). This determines which MPI ABI is supported and whether passwordless SSH or integrated launcher arguments are required.