A performance assessment system is needed to allow for centralized collection of information that describe the overall performance of the system.

The idea is to use the central ThreadCollector (which is implemented by the FawkesThreadManager) to get timing information of all threads. This information can then be represented for example by a Gantt chart. This should give clear information on the overall performance allowing to identify dead spots in the code and crowded regions (too many threads active at the same time).

An idea is to add functionality to a thread to collect this information by itself in a memory region it shares with the thread collector. The thread collector can then access this memory region just from time to time to have minimum interference of the measuring and the normal loops of the threads. If we assume to collect only the times for the begin and end of a Thread::loop() run 16 bytes are required per loop run.
If we assume that for a whole performance assessment we want to be able to record at least 20 minutes and we further assume that the thread runs at 100Hz this makes 20 * 60 * 100 * 16 = 1920000 Bytes ~ 1.84 MB. If we further assume that we want to assess 50 threads at the same time (currently if three vision processes run we have about 15 threads) we need about 100 MB of memory for collecting the information. Given that we have 2 GB and hardly used more than 10 MB this gives plenty of room.
The structures are likely to change to collect more information but additionally most threads will only run at 30 Hz and thus the rough calculation should still hold.

Using a memory region and not neither calling a handler in the a performance assessment class to collect the information nor writing to a file should give minimum influence of the measurement on the actual software behavior.