H5PY键读取速度很慢

Question

I've created a dataset with 1000 groups, each with 1300 uint8 arrays of varying lengths (though each one has a fixed size). Keys are strings of ~10 characters. I'm not trying to do anything tricky while saving (no chunking, compression etc - the data is already compressed).

Iterating over all keys is extremely slow the first time I run a script, though speeds up significantly the second time (same script, different process called later), so I suspect there is some caching involved somehow. After a while performance resets to the terrible level until I've waited it out again.

Is there a way to store the data to alleviate this problem? Or can I read it differently somehow?

Simplified code to save

with h5py.File('my_dataset.hdf5', 'w') as fp:
    for k0 in keys0:
        group = fp.create_group(k0)
        for k1, v1 in get_items(k0):
            group.create_dataset(k1, data=np.array(v1, dtype=np.uint8))

Simplified key accessing code:

with h5py.File('my_dataset.hdf5', 'r') as fp:
    keys0 = fp.keys()
    for k0 in keys0:
        group = fp[k0]
        n += len(tuple(group.keys())

If I track the progress of this script during a 'slow phase', it takes almost a second for each iteration. However, if I kill it after, say, 100 steps, then the next time I run the script the first 100 steps take < 1sec to run total, then performance drops back to a crawl.

Answer 1

While I'm still unsure why this is still slow, I've found a workaround: merge each sub-group into a single dataset

with h5py.File('my_dataset.hdf5', 'w') as fp:
    for k0 in keys0:
        subkeys = get_subkeys(k0)
        nk = len(subkeys)
        data = fp.create_dataset(
            'data', shape=(nk,),
             dtype=h5py.special_dtype(vlen=np.dtype(np.uint8)))
        keys = fp.create_dataset('keys', shape=(nk,), dtype='S32')
        for i, (k1, v1) in enumerate(get_items(k0)):
            keys[i] = k1
            data[i] = v1