issue_comments: 777602757

I've been playing around with this some more, and found the performance to be much better using a process-heavy dask scheduler. For example: ```python import xarray as xr import numpy as np import time import dask ​ def exponential(x, a, xc): return np.exp((x - xc) / a) ​ x = np.arange(-5, 5, 0.001) t = np.arange(-5, 5, 0.01) X, T = np.meshgrid(x, t) Z1 = np.random.uniform(low=-5, high=5, size=X.shape) Z2 = exponential(Z1, 3, X) + np.random.normal(scale=0.1, size=Z1.shape) ​ ds = xr.Dataset( data_vars=dict(var1=(["t", "x"], Z1), var2=(["t", "x"], Z2)), coords={"t": t, "x": x}, ) ​ ds = ds.chunk({'x':10}) ​ def test_fit(): start = time.time() fit = ds.var2.curvefit( coords=ds.var1, func=exponential, reduce_dim="t", ).compute() print(f'Fitting time: {time.time() - start:.2f}s')

with dask.config.set(scheduler='threads'): test_fit() with dask.config.set(scheduler='processes'): test_fit() with dask.distributed.Client() as client: test_fit() with dask.distributed.Client(n_workers=8, threads_per_worker=1) as client: test_fit() On my 8-core machine, takes: Fitting time: 8.32s Fitting time: 2.71s Fitting time: 4.40s Fitting time: 3.43s ``` According to this the underlying scipy routines should be thread safe.