Doubts regarding parallelization

[Lautaro Alvear]

Hi Dedalus users, I have a working code that simulates 3D compressible magnetohydrodynamics in a star, but the code runs extremely slow without parallellization to solve the timescales of interest (~100-1000 time units with a timestep of dt ~ 1e-5) with the current test resolution (16,48,96). So I have been trying to parallelize the code with up to 48 cores, which has been succesfull only if I use a mesh of shape [1,n] with n the number of cores. If I try to use a different value than 1 for the first value of the mesh, the code raises exception at runtime, which to me would signify that the dynamics are changing with the mesh, which as I understand, should not be the case, unless, the shape [1,n] is forcing axial symmetry. I wanted to ask if this is the case, and if not, if someone has encountered similar issues when parallelizing.

Best regards,

Lautaro.

[Daniel Lecoanet]

Hi Lautaro,

Sounds like you are using the code in some suboptimal way. From what you are describing I cannot tell what is going wrong. There are multiple ways to formulate the compressible equations, some of which are much more efficient than others. I would recommend looking at this reference for formulations of the compressible equations (https://ui.adsabs.harvard.edu/abs/2014ApJ...797...94L/abstract) and this paper for 3D compressible simulations of a star (https://ui.adsabs.harvard.edu/abs/2023NatAs...7.1228A). In the latter case, the simulations were run on 16384 cores without any problems. Otherwise, I’ll note we have tested dedalus on up to 10^5 cores and found very good scaling.

Daniel

--
You received this message because you are subscribed to the Google Groups "Dedalus Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dedalus-user...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/dedalus-users/c3d49e31-1cca-4355-9ded-97f9868799fan%40googlegroups.com.