Home > Uncategorized > Visualising Your Process Network

Visualising Your Process Network

One advantage of CHP’s process-oriented style of programming is that it lends itself well to visualisation. I don’t make as much use of it on this site as I should, but some previous posts show diagrams with processes (as nodes) joined by channels (as edges). I already have some tracing support in CHP, and I had wondered if this could provide a quick and dirty way to graph a process network. This post shows the results of adding a little support to CHP for process visualisation, which I’m now debating whether or not to keep.

It turned out to be slightly better to add a new trace type (the process-graph trace) to CHP than to build a graph from the existing structural traces. So from the user’s point of view, all that would be added to the library was a new trace type; no changes to an existing program would be required other than switching tracing on and using the process-graph as the trace type. (I also added a labelling function for processes, that works with both structural and process-graph traces.)

Hooking the process-graph trace into the existing tracing infrastructure was so straightforward that it’s not really worth writing about. Unexpectedly, the most difficult part of making this functionality useful was finding something that can draw the process networks well. Process networks tend to be hierarchical, in that each process (node) may contain a sub-network, composed to any depth. These processes are connected via edges. My first inclination for graph visualisation was to reach for graphviz. So I wrote some code to spit out the graphviz file. Graphviz has a dot tool for drawing graphs in lines with rank. On a simple benchmark called numbers, it does quite well:

For larger examples, or those with cycles, its rank-basis can be awkward. Here’s the dining philosophers, rendered in a slightly unusual manner:

Most of the other graphviz tools can’t deal with hierarchical graphs — only the fdp tool can. This seemed promising; a force-directed approach that supported sub-graphs. Here’s the numbers example from above:

Clumsy. And here’s the dining philosophers:

Eugh! This is a tiny bit better when you liberate it from its subgraph (at least, it might be halfway decent if not for the node labels):

But there is a clear solution for the dining philosophers (a star topology with the security guard at its centre), so if fdp makes a hash of this graph, I don’t hold out much hope for more complex process networks. I don’t want to write my own tool for this, but I may if I have to (I’d actually quite like an interactive viewer so that you can drag things around like you can in prefuse, and perhaps involve animation of the network). Any suggestions for beating fdp into shape, or for alternative tools would be most welcome!

Observed Occurrences Only

I said that this technique was quick and dirty. The main flaw (or is it a feature?) is that only the observationally determinable process network is graphed. If you pass a channel to two processes who never communicate on it, that link won’t show up in the graph. The graph is a trace: a record of what happened, not a blueprint of what might have happened. This is not an invalid approach, but I expect it is not what most people want: most would like to see their process network drawn as it is in their program design with all channels shown, not just those that were used. But that tool is more complicated, and will have to wait for another day. (You can extend this hack to do something more like that if you use an annotation and generic programming to capture all processes’ parameters and look through them for channel ends…)

Change and Movement Over Time

The fallacy of drawing the process network in a single image is that it is not static for the duration of the program. Processes can start and end, come and go. Not only that, but CHP channels can be freely communicated down other channels, meaning that programs can dynamically rewire their topology at run-time (I should write some posts about that at some point). So really a proper tool should include support for visualising the process network as it changes over time. But again, that will have to wait for another day.

About these ads
Categories: Uncategorized Tags:
  1. May 11, 2010 at 2:44 pm

    Have you considered using the ‘circo’ layout? I find it generally does a better job of clarifying the edges in a graph diagram than the other graphviz layouts.

    • May 11, 2010 at 2:54 pm

      Circo does do better at the layout of a non-nested graph, but it doesn’t support nested sub-graphs with visible borders like dot and fdp do. My above examples are perhaps a bit too simple to show the power of the nested structure of CHP programs, but that was something that would be very useful for real examples. Though interactive hiding/zooming of the nested networks would be particularly good — and that would need an interactive tool rather than just a static image.

  2. Stephen Tetley
    May 11, 2010 at 7:26 pm

    Hi Neil

    If you can live with a chain of tools – LaTex / dvips to make a PostScript file, GraphicsMagik to make a JPG or whatever, TikZ is possibly the finest tool there is.

    Check the examples page to see if there is a readymade style close enough for your purposes, otherwise the manual is pretty comprehensive for describing how to make network diagrams.

    http://www.texample.net/tikz/examples/

  3. Matthias Neubauer
    May 11, 2010 at 7:43 pm

    Have a look at yEd. It does a pretty good job of rendering hierarchical graphs.

  1. No trackbacks yet.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

Follow

Get every new post delivered to your Inbox.

%d bloggers like this: