One other important limitation I found and I forgot to mention. At the start of the execution of the flowgraph, there is a phenomenon which I interpret like this: before all downstream buffers are filled, there is no constraint on the speed of execution of the work() function in the source block, probably due to lack of back-pressure?
So this resulted in lots of timeslot samples created in a very short amount of time (microseconds?) that push the initial latency to 10 seconds or more. So I've constrained the execution of the work() function at the startup which adds those 30 msec which are not caused by the flowgraph. Unfortunately I could not figure out how to lift that limitation once back-pressure takes effect (all downstream buffers are full). I hope that makes sense. Adrian
