Walberla: Flag field should not be ghost communicated at every time step

[RudolfWeeber]

In benchmarks, Espresso-dev with a pure LB fluid has worse scaleability than 4.2.1 but better performacne on low numbers of cores.

Possible cause: the flag_field (for boundaries) is part of the m_full_communication pack info whic is used for ghost communication after every time step. This is, to my understanding not necessary. The flag field only changes, when cells are marked as fluid/boudnary.

Possible solution:
have two pack infos,

• one m_pdf_vel_force_communication for pdf, last applied force and velocity, to be run once per time step
• one m_flag_field_communicaiton.

The first should be called in LBWalberlaImpl::ghost_communicaiton(), the second only when the fluid/boudnary state of a cell changes.

BTW: the pull scheme integraiton already calls ghost_communicaiotn(), the push scheme directly calls the call operator of the pack info. This can be changed to ghost_communicaiton() as qwkk,

Further options:

• not ghsot communicating the velocity field and instead recalculate it in the ghost layers from pdf and last applied force. this will be more work to implement, though.

[jngrad]

Quick benchmark on 8 cores: removing the flag field from the ghost communicator speeds up simulations by 2%.