Make it possible to override the default chunking behavior when loading to dask arrays. For PIConGPU the default creates a chunk for each GPU with its local domain. When using static load balancing with gridDist in PIConGPU, this results in very uneven chunks. It looks like, for many operations, it makes sense to chunk along the outermost direction only (reading contiguous domains is faster). With this PR this can be achieved with .to_daks_array(chunks={0 : 'auto', 1: -1, 2: -1})
Make it possible to override the default chunking behavior when loading to
dask
arrays. ForPIConGPU
the default creates a chunk for each GPU with its local domain. When using static load balancing withgridDist
inPIConGPU
, this results in very uneven chunks. It looks like, for many operations, it makes sense to chunk along the outermost direction only (reading contiguous domains is faster). With this PR this can be achieved with.to_daks_array(chunks={0 : 'auto', 1: -1, 2: -1})