Dropping trials with rtmixer backend

[tomstoll]

We've been dropping trials occasionally when delivering our sounds and triggers through the sound card (using the rtmixer backend). No sounds or triggers are sent. A recent experiment we were piloting that had tens of thousands of 100 ms long trials made the problem more obvious. Less than 1% of trials are dropped and it's not a big issue for that experiment, but losing a trial in experiments with fewer trials could be pretty annoying. We ran another test with 540 one second long trials and we dropped two trials.

We were using the rtmixer backend to play sounds and triggers through the sound card. We haven't tried the pyglet backend but I could see if it'll work with our setup. In a test with the sounddevice backend I'm working on (#494), we dropped zero trials (out of 7200, all one second long).

Not sure where the problem is exactly, maybe something to do with loading the buffer if the computer is under load? I can try some type of stress test while running an experiment and see if that forces trials to drop (with any backend).

[larsoner]

I think I've had that problem with rtmixer in the past just in a simple loop playing a sound or something (it was a couple of years ago and I don't remember much else, just rang a bell). I never got to look into it so not sure if it's a problem with rtmixer (most likely) or PortAudio itself or what.

[rkmaddox]

I'll just add that we have seen this across several computers and sound cards (all RME, though) over the last few years. Just haven't gotten around to dealing with it yet.

[tomstoll]

Just adding for reference that we do still drop some trials with the sounddevice backend. Still not clear what happens, since I typically still see the triggers. Since we have SOUND_CARD_TRIGGER_ID_AFTER_ONSET=True, this seems to indicate that the audio was sent (at least the start of it) before cutting out.

With gapless playback, I also see in the EEG software that the triggers shift later after a dropped trial (i.e., the dropped trial lasts longer than a normal trial).

Not sure what this means, but maybe one day it'll be useful in solving this.