RedistributeCPU with 1 Core 1 Thread 1 Box

[ax3l]

RedistributeCPU should be a no-OP if run with 1 MPI rank and 1 OpenMP thread.

In a benchmarked of Warp vs WarpX shown by @dpgrote today, it does in fact still take ~20% of the time of the 2D simulation loop in WarpX.

Am I overlooking something? :)

[atmyers]

Are there still multiple Boxes and tiles in the simulation?

[dpgrote]

The simulation had one box.

[atmyers]

If particles.do_tiling = 1, which I believe is the default in WarpX when compiling with OpenMP, there will be multiple tiles even on that one Box, so Redistribute will still have some work to do. We will try to optimize this, but to have it truly be a no-op, you have to set particles.do_tiling = 0.

[dpgrote]

In my test case, if I set particles.do_tiling = 0, it doesn't affect the timings. RedistributeCPU takes the same amount of time.

[atmyers]

Can you post the inputs here so I can run this?

[dpgrote]

This is the same input I used here BLAST-WarpX/warpx#6487