Profile and reduce simulation overhead

[gizatt]

Opening an issue to start to collect some measurements in terms of possible causes of inefficiency in the sim (esp. on our end with ROS interface things).

Here is a callgrind file from a sim run; view with e.g. KCacheGrind.

Some observations so far (preliminary, and not totally confident... this code includes lots of cycles that confuse the profiler...)

• MBP simulation itself seems to take a majority (but not a huge majority) of the time, but with more cameras, rendering might take the lead.
• The slowest part of the RosRgbdCameraPublisher is its loop copying the depth image into the ROS message, but it takes about 1/5 the time as the rendering itself. (So there's lots room for improvement, especially as it's a relatively straightforward simple memory copy, but it's not egregious as we worried.)
• The RViz Scene Graph viz is definitely doing something bad -- the IM ApplyChanges() method is taking something like 5% of all runtime, and seems to be on the same order of magnitude (maybe 2x-4x less?) than the RGBD rendering.

But again these are viewed through the lens of profiling code that gets called from a really complicated cyclic call graph (I think, since the systems framework does many evaluations on-demand by dependency tracking, e.g. so the RGBD camera publisher system actually gets "credit" for the RGBD rendering). I'll poke a bit more at doing this profiling in other ways (via abalation, and via some hooks into the System framework...)

[peteflorence]

🥇