Improve performances in fetching data

[gmaze]

Add an asynchrone open_mfdataset to the httpstore
and much more ...

Close #27 #16 #51

[gmaze]

Here are some benchmarking results for the time of data fetching using the erddap and a large box.
Box limit varies to increase dataset to fetch in size.

The following code was repeated for box depth of 0-50m and 2018 alone and then 0-50m, 0-100m, 0-200m, 0-300m and 2018/2019.
The horizontal domain size is 40 deg in longitude and 20 deg in latitude.
I use small random variations around the longitude boundaries to avoid the erddap server to cache requests.
I also use the expert mode to limit the amount of postprocessing on the client side.
Note the new fetcher options parallel=True, chunks='auto', box_maxsize=[10, 10, 50].
We run the same request n=5 times to get an ensemble approach and more stable metrics.

for run in range(5):
    start_time = time.time()
    large_box = [-70+np.random.random_sample(1)[0], -30+np.random.random_sample(1)[0], 20, 40, 0, 200, '2018-01-01', '2020-01-01']
    fetcher = ArgoDataFetcher(mode='expert', parallel=True, chunks='auto', box_maxsize=[10, 10, 50]).region(large_box)
    ds = fetcher.to_xarray()
    par_bench.append({'ETIM':time.time() - start_time, 'NPTS':np.max(ds['N_POINTS'].values), 
'MBYT': ds.nbytes/1e6, 'CHUNKS': fetcher.fetcher.chunks, 'NREQ': len(fetcher.fetcher.urls)})
    print("Run #%i" % run, par_bench[-1])

ps: The request fails for 0-400m and 0-800m boxes.

Results:

[quai20]

I'll continue testing, but code is ok to me, and based on docs everything works fine.

[gmaze]

ok, thanks @quai20 for the review
do you think, this is ready for a merge ?

[quai20]

ok, thanks @quai20 for the review
do you think, this is ready for a merge ?

ok to me !