distob 0.3.3

Distributed computing made easier, using remote objects

Distributed computing made easier, using remote objects

N.B. this is a development pre-release: still a lot left to be done

Overview

Distob will take your existing python objects, or a sequence of objects, and scatter them onto many IPython parallel engines, which may be running on a single computer or on a cluster.

In place of the original objects, proxy objects are kept on the client computer that provide the same interface as the original objects. You can continue to use these as if the objects were still local. All methods are passed through to the remote objects, where computation is done.

In particular, sending numpy arrays to the cluster is supported.

A numpy array can also be scattered across the cluster, along a particular axis. Operations on the array can then be automatically done in parallel (either using ufuncs, or by using vectorize() below)

Note: numpy with __numpy_ufunc__ feature enabled (not yet released) is required to support distributed array arithmetic and distributed ufuncs. You can get numpy with this experimental feature enabled here: https://github.com/mattja/numpy/archive/master.zip

Distob is an object layer built on top of ipyparallel, so it will make use of your default IPython parallel profile. This allows different cluster architectures, local CPUs, SSH nodes, PBS, Amazon EC2, etc.

functions

scatter(obj) Distribute any object (or list of objects) to remote iPython engines, return a proxy.

gather(obj) Fetch back a distributed object (or list), making it local again.

vectorize(f) Turn an ordinary function (that takes a single object or array) into one that acts in parallel on a scattered list or array. apply(f, obj) is the same as vectorize(f)(obj)

distributed numpy arrays

scatter(a, axis=2) Distribute a single numpy array along axis 2, returning a DistArray.

Arithmetic operations can freely mix ordinary arrays with the new array types.

Normal numpy ufuncs can also be used on the distributed arrays.

Arithmetic and ufunc computations will automatically be routed to an engine, or executed in parallel on several engines, depending on where the data is. (requires numpy with the __numpy_ufunc__ feature enabled)

concatenate, vstack, hstack, dstack, expand_dims, transpose, rollaxis, split, vsplit, hsplit, dsplit, broadcast_arrays:

These work like the numpy functions of the same name. But these can be used with a mix of ordinary ndarrays, RemoteArrays and DistArrays, performing array structural changes while keeping the actual data distributed across multiple engines.

For example, stacking several RemoteArrays gives a DistArray, without needing to move data.

The distributed arrays so far support basic indexing, slices and advanced integer indexing.

classes

RemoteArray proxy object representing a remote numpy ndarray

DistArray a single ndarray distributed across multiple engines

Remote base class, used when auto-creating Remote* proxy classes

@proxy_methods(base) class decorator for auto-creating Remote* proxy classes

ObjectHub dict interface giving refs to all distributed objects cluster-wide

ObjectEngine dict holding the distributed objects of a single IPython engine

Ref reference to a (possibly remote) object

attributes

engine: the ObjectEngine instance on each host (ObjectHub on the client)

TODO

• Allow assignment to slices of remote arrays.

• Properly implement caching of remote method results.

• Auto-creation of proxy classes at runtime (depends uqfoundation/dill#58)

• For ufunc execution, still need to implement reduce, accumulate, reduceat, outer, at methods.

• Make proxy classes more robust, adapting wrapt (pypi.python.org/pypi/wrapt)

Thanks

Incorporates pylru.py by Jay Hutchinson, http://github.com/jlhutch/pylru

ipyparallel interactive parallel computing: https://ipyparallel.readthedocs.org/

dill by Mike McKerns for object serialization, see: http://trac.mystic.cacr.caltech.edu/project/pathos