Skip to content

Files

Latest commit

author
Heonjin
Sep 9, 2020
570bcfd · Sep 9, 2020

History

History
64 lines (47 loc) · 4.94 KB

File metadata and controls

64 lines (47 loc) · 4.94 KB

Array storage benchmark

Compare the storage speed, retrieval speed and file size for various methods of storing 2D numpy arrays.

Hardware etc

The results here are obtained on a normal desktop PC that's several years old and running Ubuntu and has a SSD for storage. You can easily run the benchmarks on your own PC to get more relevant results. You can also apply it to your own data.

Methods

Name Description Fast Small^ Portability Ease of use Human-readable Flexible% Notes
Csv~ comma separated value ☐ ☐ ☐ ☐ ☐ ☐ ☒ ☒ ☒ ☒ ☒ ☒ ☒ ☒ ☒ ☐ only 2D
JSON~ js object notation ☐ ☐ ☐ ☐ ☐ ☐ ☒ ☒ ☐ ☒ ☒ ☐ ++ ☒ ☐ ☒ ☒ any dim, unequal rows
b64Enc base 64 encoding ☒ ☒ ☒ ☒ ☐ ☐ ☒ ☒ ☐ ☒ ☒ ☐ ☐ ☐ ☐ ☐ more network, not files
JsonTricks json-tricks compact ☒ ☒ ☐ ☒ ☒ ☐ ☒ ☐ ☐ ☒ ☒ ☒ + ☐ ☐ ☒ ☒ many types beyond numpy
MsgPack Binary version of json ☒ ☒ ☒ ☒ ☒ ☐ ☒ ☐ ☐ ☒ ☒ ☐ + ☐ ☐ ☒ ☐  
Pickle~ python pickle ☒ ☒ ☐ ☐ ☐ ☐ ☐ ☐ ☐ ☒ ☒ ☒ ☐ ☐ ☒ ☒ any obj, not backw. comp
Binary~ pure raw data ☒ ☒ ☒ ☒ ☒ ☐ ☒ ☒ ☒ ☒ ☐ ☐ ☐ ☐ ☐ ☐ dim & type separately
NPY numpy .npy (no pickle) ☒ ☒ ☒ ☒ ☒ ☐ ☒ ☐ ☐ ☒ ☒ ☒ ☐ ☐ ☒ ☐ with pickle mode OFF
NPYCompr numpy .npz ☒ ☒ ☒ ☒ ☒ ☒ ☒ ☐ ☐ ☒ ☒ ☒ ☐ ☐ ☒ ☐ multiple matrices
PNG encoded as png image ☒ ☒ ☐ ☒ ☒ ☒ ☐ ☐ ☐ ☐ ☐ ☐ ++ ☐ ☐ ☐ ☐ only 2D; for fun but works
FortUnf fortran unformatted ☒ ☒ ☒ ☒ ☒ ☐ ☒ ☐ ☐ ☒ ☐ ☐ + ☐ ☐ ☒ ☐ often compiler dependent
MatFile Matlab .mat file ☒ ☒ ☒ ☒ ☒ ☐ ☒ ☒ ☐ ☒ ☒ ☒ + ☐ ☐ ☒ ☐ multiple matrices
  • ^ Two checks if it's small for dense data, three checks if also for sparse. All gzipped results are small for sparse data.
  • % E.g. easily supports 3D or higher arrays, unequal columns, inhomogeneous type columns...
  • ~ Also tested with gzip, stats refer to non-gzipped. Gzipped is always much slower to write, a bit slower to read, for text formats it's at least 50% smaller.
    • Rating refers to using a semi-popular package (probably scipy), as opposed to only python and numpy.
  • ++ Very easy (☒☒☒) with an unpopular and/or dedicated package, but the rating refers to only python and numpy.

You can install all dependencies using pip install -r requirements.pip. csv and NPY were done with numpy; json and compact json (JsonTricks) were done with pyjson_tricks; png was done with imgarray; fortran unformatted and matlab were done with scipy; pickle, base64 and gzipping were done with python built-ins. HDF5 uses h5py (not finished, see issue4). MessagePack uses msgpack-numpy. Seaborn is needed for plotting. You can install all dependencies using pip install requirements.pip

Results

Dense random matrix

/result/bm_random_integer.png

More methods

Pull requests with other methods (serious or otherwise) are welcome! There might be some ideas in the issue tracker.