Simple two-process client/server plotting with the following features:
- Plotting does not slow down calculations.
- User maintains control of the calculation (i.e. KeyboardInterrupts work).
- Auto-launch client from the server.
- The user defines the plotting code in a (picklable) class, and provides the class to the Server() constructor. The Server passes the pickled Plotter to the client when the client process is launched in a new python interpreter. The python interpreter command line options for the client are passed to the Server constructor at the time of instantiation. Default values are: (ex. ["ipython", "--pylab=osx", "-c"])
Prototype for asynchronous plotting with separate processes using sockets
- Comprehensive testing
- Logging (print debug messages in debug mode for checking network problems)
- Profile performance.
- Configuration (hostname, port specification etc.)
I think that these have been dealt with, but they need testing:
- socket buffer overflow (recv)
- cleanup thread and socket command line termination
- error handling. ex. when a client disconnects then server listen continues, network errors
- multiple clients (plotting)
Other Design Approaches
The simplest approach is a multi-thread approach. In principle, one can run the computations in the main thread and plotting in a separate thread. This solution is sketched in thread.py but fails with most matplotlib backends due to their requirement of running in the main thread. A quick work-around is to run the computation in a secondary thread, but this precludes the user being able to interrupt the computation.
A nice feature of the python GIL is that one can be fairly confident about sharing data (a careful solution would require locks etc.)
This same solution should work with multiprocessing, but this fails on my development platform (Mac OS X 10.5) with the following error:
The process has forked and you cannot use this CoreFoundation functionality safely. You MUST exec(). Break on __THE_PROCESS_HAS_FORKED_AND_YOU_CANNOT_USE_THIS_COREFOUNDATION_FUNCTIONALITY___YOU_MUST_EXEC__() to debug.
It seems that the most robust solution is to have the calculation and plotters run in completely separate processes. This has an added benefit:
- User can plot remotely.
One issue that needs to be addressed here (and in the multiprocessing solution) is the copying of data. One common use-case is that the plotter may be slower than the computation. Thus, intermediate data may be discarded and should not be sent across the network.
It seems that one should be able to use IPython to do this, but I have not found a simple way to do this yet.