[visionegg] Re: time and OS es


On Mar 28, 2005, at 4:44 PM, Christophe Pallier wrote:

Based on non casual observations, I can tell you it is too frequent , at least on my hardware and for some experiments.
(My PC is a 2 Ghz pentium 4/Nvidia Gforce4 and with all the (unpatched) linux kernels I have tried over the years, I get several latencies above 30 msec every minute. To my dismay, VisionEgg demos almost always skip a few frames even in realtime mode. Maybe I should change my computer...).

OK, let's say it is 7 every minute. Now, anyone want to work out the odds of missing one of the couple hundred critical milliseconds out of the several thousand in that minute? And, you can always toss the rare time that occurs by actually recording that it did.

Yet, the results are obsviously dependent on hardware, version of kernel, and the load while testing (for example, the BRMIC paper you cite, if I remember correctly, ran the test on a machine which did not do any stimulus presentation. I was not impressed.).

you and me. But, I would have hoped that it stimulated a follow up that did it right. No such luck...

But let me just give some real life cases where occasional latencies could be a problem:

- The stimulation PC must not miss the 5 millisec-long TTL signals sent from a MRI scanner to synchronize trials with brain scans.

Now, this is a potential problem in any event. Note that in this event missing the critical msec that the signal occurred wouldn't be so bad (100 msec slop on the detection would be pretty meaningless in most fMRI and 30msec would be fine) except that there is no way to make up for it. I suggest that you always latch these outputs. If the scanner triggers it goes high and stays high until the computer resets. That's really some dead simple electronics.

- In ERP experiments, subliminal stimuli are displayed for 2/3 video frames, several hundreds of times in 10 minutes runs. It is very important that the subject does not see consciously any of the subliminal images because it could change his strategy. So one would like to have frame-by-frame accuracy for several minutes.

agreed, problematic if they see one in that instance. However, in most studies of these type it is OK if they see the occasional sub stimulus (and there is really never any guarantee that they don't). In that case, the odds of missing one of the critical frames is probably perfectly acceptable (and a recordable error).

Which brings out one of the points I made but didn't emphasize. Many experimenters do not record the error of the measurement. A long tradition of canned packages that just give you back what you put in has really promoted this. I'm glad some of the newer ones like Experiment Builder (SRI) and Presenter(?) record what happened.

Good experimenters must know the limits of their tools, and how to deal with them.
I have encoutered more than one who tells you that nowadays machines are fast enough and who does no check the timing. 

agreed

======================================
The Vision Egg mailing list
Archives: http://www.freelists.org/archives/visionegg
Website: http://www.visionegg.org/mailinglist.html

Other related posts: