[OVS]

Hello,

I am trying to solve a steady RANS airflow around a 3D Wing with 5 million nodes (15 M elts, unstructured cgns).

I have acces to a cluster, however it seems that after a given amount of cores (192) the solver has problems in the MPI communications and memory management department. It usually exits with a:

mpirun noticed that process rank XX with PID YYYYY on node ab1234 exited on signal 11 (Segmentation fault).

Talking with the technicians of the cluster they tell me that at some point SU2 violates its allocated memory for a given thread (which crashes the calculation). It seems to me that SU2 loads all the work into the master node either while reading (at the begining, if restarting) or writing solutions (be it the restart file or - mainly - when generating the flow.dat file).

I'm running on a cluster with 24 ppn and 32GB per node - swiching from 8 to 12 nodes to find the one configuration that *might* work -.

A problem of this size shouldn't be difficult to solve for this type of cluster... I asked around and a colleague solved a DNS simulation with around 200M nodes in roughly 100h with 144 cores (2 years ago), yet it takes me roughly 60h for a 5M node compressible steady RANS-SA simulation with 192 cores... which sometimes crashes because of single thread memory issues.

Has someone had problems scaling SU2 (using several processors)? How many processors have you managed to use at max? Have you done an efficiency test (to find the point where you no longer gain time with parallel calculations)? What about memory usage?

Thanks,
Oliver.