imclab/pyrallel

Experimental parallel data analysis toolkit.

Pyrallel - Parallel Data Analytics in Python

Overview: experimental project to investigate distributed computation patterns for machine learning and other semi-interactive data analytics tasks.

Scope:

• focus on small to medium dataset that fits in memory on a small (10+ nodes) to medium cluster (100+ nodes).

• focus on small to medium data (with data locality when possible).

• focus on CPU bound tasks (e.g. training Random Forests) while trying to limit disk / network access to a minimum.

• do not focus on HA / Fault Tolerance (yet).

• do not try to invent new set of high level programming abstractions (yet): use a low level programming model (IPython.parallel) to finely control the cluster elements and messages transfered and help identify what are the practical underlying constraints in distributed machine learning setting.

Disclaimer: the public API of this library will probably not be stable soon as the current goal of this project is to experiment.

Dependencies

The usual suspects: Python 2.7, NumPy, SciPy.

Fetch the development version (master branch) from:

• https://github.com/ipython/ipython

• https://github.com/scikit-learn/scikit-learn

StarCluster develop branch and its IPCluster plugin is also required to easily startup a bunch of nodes with IPython.parallel setup.

Patterns currently under investigation

• Asynchronous & randomized hyper-parameters search (a.k.a. Randomized Grid Search) for machine learning models

• Share numerical arrays efficiently over the nodes and make them available to concurrently running Python processes without making copies in memory using memory-mapped files.

• Distributed Random Forests fitting.

• Ensembling heterogeneous library models.

• Parallel implementation of online averaged models using a MPI AllReduce, for instance using MiniBatchKMeans on partitioned data.

See the content of the examples/ folder for more details.

License

MIT

History

This project started at the PyCon 2012 PyData sprint as a set of proof of concept IPython.parallel scripts.