So far, everything we've run was run on one CPU, sequentially—with the exception of some ML models and transformations, which support the number of jobs (parallel executors); for example, cKDTree supports multiprocessing, if needed.
The caveat here is the overhead—in order to run a multicore process, a lot of additional memory needs to be allocated and data needs to be copied; it is essentially a fixed cost. Because of that, most of the tasks we ran wouldn't benefit from multiple cores, except for cases where data is very large and computations are fairly parallelized. On the flip side, once we run a task on multiple cores, spreading it across multiple machines is simple.
While the most typical task for Dask to deal with is heavy computation ...
