Latency variation

Question

I have a problem using my raspberry pi 2 for applications that need low latency.

I wrote a c++ GPIO routine that reads a high resolution quadrature ttl encoder. It then updates a counter and adds it to a value being read by i2c channel from another thread. Then another GPIO pins create a quadrature ttl encoder output. To not lose any pulses I need at least a loop with 10 usec cycle time for encoder reading thread.

The program cycle time is less than 3 usec in 99.9% of the times. Only one out of 1000 time the cycle time becomes more than 2000 usec.

I think the os is running so many unnecessary tasks that I don't need. For instance I don't need graphics, desktop, networking, Bluetooth, USB or even keyboard or monitor. I tried to stop all the running services with no luck. Any help would be much appreciated.

Answer 1

I need at least a loop with 10 usec cycle time

You won't be able to get that consistently with a userspace application on a non realtime OS. A multi-tasking OS simply does not work that way; even if you minimize the number of processes, you cannot guarantee sub millisecond latency -- in my experience (which is admittedly limited in this regard), the lowest granularity you can get consistently with the linux kernel is 10 ms. Although you might get much lower than that regularly, eventually it will be off.

I don't know whether you can do better than that in kernelspace. You might be able to work out a way to predictably do it for short time periods (like, 10 ms) but at some point you must be at the mercy of the scheduler.

You can build the kernel with options to minimize latency but even those refer to "the milliseconds range". There are true realtime linux kernels but they are a separate project and so combining them with the pi kernel, also a spin-off, is daunting. This Q&A discusses some possibilities.

Not really mentioned there is RISCOS -- or rather, it's mentioned in a comment but then someone points out it is not a realtime OS. Which is true, but it does use cooperative multi-tasking (CMT). I have never used RISCOS but I did have to read a bit about CMT systems in school (they were common up into the 80's) and I think it is worth at least doing a bit of research about. Most likely, you cannot safely run a process without ceding control regularly, but you could guarantee that it will not be interrupted for short periods.