Now everything in the remote memory footprint sits inside an object called an 'Endpoint'. The interfaces to set up and hold fields, etc. are generic, and are initialized as callables in the remote namespace, so they should be able to be adapted to different data backends (Fields objects, HDF5, etc.).
SuperReference acts like an IPython.parallel.Reference instance specialized for function calls. If it's scheduled by a call to 'apply' on a Load Balanced View, it will raise UnmetDependency errors until it's scheduled on a worker that is allowed by its 'rank' attribute. This can therefore point at functions that are known to be defined on particular workers.
Endpoint is a prototype layout for an object to hold multiple Problems, Fields, etc. on the remote workers. The idea is to clean up the remote namespace and reduce the use of globals().
DataWrappers.py contains the CommonReducer class, which enables pass-through math operations and function calls on dictionaries.
Parallel.py contains the job scheduler, remote interface (w/ MPI support) and several helper functions.
Brendan Smithyman