[sugar] more on sugar performance

[Dan Williams]

On Thu, 2006-09-21 at 21:49 +0800, Ivan Krstić wrote:
> Dan Williams wrote:
> > You're not seriously suggesting we write our own general purpose IPC
> > system, are you? 
> 
> No, I'm not. That's complete madness ;)
> 
> I do want to look at D-BUS more carefully. Maybe there are ways to tweak
> it that would make it more pleasant for the way we use it.

Quite certainly; we also may not be using it correctly.  For one, the
Python bindings try to be _quite_ clever.  They do introspection calls
at least once for every method call, to determine the arguments for the
method call.  So in some instances you end up dong 2 method calls (one
for introspect, one for the real call).  That adds latency.

Also, if we want to shove gobs of data down the pipe, we may wish to use
private Point-to-Point connections rather than grabbing a connection to
the bus.  It's fairly known what the limitations of the bus-connected
method is, versus the PtP method, and what the tradeoffs are.

I don't think that anyone has spent a huge amount of time profiling the
Python D-Bus bindings, certainly a lot less than any time spent
profiling libdbus.  I think there's a fair bit of room for improvement
here.

Dan