在图形并行计算中处理相互依赖的文件

Dealing with interdependent files in graph-parallel computation

我正在尝试并行化以下代码(MCVE) by creating a task graph using dask.delayed (or by implementing a computational graph 我自己):

os.chdir('./kitchen1')
write_dough()   # writes file ./dough
write_topping() # writes file ./topping
write_pizza()   # requires ./dough and ./topping; writes ./pizza

我看到 2 个困难:

  1. write_dough 没有 return 任何东西。 z=x+y让变量之间的依赖关系一目了然;这不是。 Dask doesn’t recommend relying on side effects。有惯用的解决方案吗?
  2. os.chdir。如何将其合并到计算图中?
  3. 我不关心并行化文件 IO、性能等

这是我目前的解决方案。它增加了复杂性,并且 './kitchen1' 无处不在,这很丑陋。优雅的解决方案是什么?

write_dough, write_topping, write_pizza = map(dask.delayed, (write_dough, write_topping, write_pizza))

dough = write_dough('./kitchen1')
topping = write_topping('./kitchen1')
pizza = write_pizza(dough, topping, './kitchen1')

我会推荐您当前显式传递依赖项的方法。