doudou
commented
6 years ago I use a composition to generate input commands for the auv_control
poll is called at the periodicity of the Syskit event loop. Its periodicity is not guaranteed.
However, the plan generation is meant to be asynchronous, leaving the event loop free to react to things (and call poll), The log you've just shown tends to prove otherwise - the 2s delay you're seeing seem to be related to the network generation since there are no timepoints related to the event loop within the 2s period. This is seems to be a Syskit performance / setup bug that needs to be investigated.
Moreover, ConstantGenerator does the generation within a thread to guarantee a better periodicity. While this may be affected by the same issue, I would suggest to refactor ConstantGenerator into a ThreadedPortWriter abstract base class where most of the thread-handling would be done, and use this as a basis for your sinusoidal generator. This would definitely improve reactivity. However, it could also be affected by the same thread scheduling bug.
joaobrittoneto
I use a composition to generate input commands for the auv_control. It happens that the auv_control was expecting to receive
cmd_in(provide byDive::Compositions::SinusoidSignalGenerators::WorldPos) with at least 1s.It happens that sometimes, the poll block of the compositions are called with a ~2s interval (16:28:51, 16:28:53, 16:28:55). Just increasing the timeout would fix the problem, but I'd like to understand what Syskit is doing in the meantime. Is there a way to guarantee the periodicity of a composition's poll?
Here is the data provided by
roby-log decode live-events.log