Performance evaluation for LBM

Dear all,

I’ve written an LBM code and I would like to evaluate it’s performances. According to various paper this is given in LUPS (i.e. lattice updates per seconds).

Does anyone know how this quantity is computed in practice?

Is it enough to run a simulation for Nt timesteps using a mesh of size NxNyNz, and compute the performace P as:

P = NxNyNz*Nt/(computational time) ?

Is this the correct way to evaluate the performances in LUPS?

If it is so, according to me this kind of evaluation may vary depending on the test case used as well on the number and type of boundary conditions employed. Is it right? As a consequence, is there any reference benchmark that you can suggest me in order to test my code performances?

Thanks a lot


your definition of the LUPS is indeed correct with P=LUPS.

For Palabos the standard test case is the 3D lid-driven cavity. You can find the results on various platforms and grid sizes here.