`mapCube` in Distributed model is many times slower

Serial: 8s

using YAXArrays, Statistics, Zarr
using DimensionalData
using Dates
axlist = (
    Dim{:time}(Date("2022-01-01"):Day(1):Date("2022-12-31")),
    Dim{:lon}(range(1, 10, length=1000)),
    Dim{:lat}(range(1, 5, length=1500)),
    Dim{:Variable}(["var1", "var2"])
    )

data = rand(axlist...)
ds = YAXArray(axlist, data)
c = ds[Variable = At("var1")] # see OpenNetCDF to get the file
c2 = setchunks(c, (365, 100, 150))

function mymean(output, xin)
    # @show "doing a mean"
    output[:] .= mean(xin)
end

indims = InDims("time")
outdims = OutDims()
@time resultcube = mapCube(mymean, c2; indims, outdims)

Parallel: 660 seconds

using Distributed
addprocs(4)
# @everywhere using Pkg
# @everywhere Pkg.activate(".")
@everywhere begin
  # using NetCDF
  using YAXArrays
  using Statistics
  using Zarr
end

@everywhere function mymean(output, xin)
    # @show "doing a mean"
    output[:] .= mean(xin)
end

indims = InDims("time")
outdims = OutDims()

# turn out to be slower? why?
@time resultcube2 = mapCube(mymean, c2; indims, outdims)

Start by reproducing the serial and Distributed examples in the issue, focusing on the mapCube entry point and the c2 chunking setup. Compare the parallel execution and data movement with the serial run; done means identifying the cause of the slowdown and documenting or correcting the behavior with evidence from the benchmark.