[TriLUG] High resolution timer calls and the kernel

[Sam Kalat]

I have a real-time computer vision project I'm working on that does some 
analysis of incoming 10 fps video from a pair of webcams.  The cycle of the 
cameras is reliable so each frame comes in 100 msec after the last.  The only 
exception is the first frame which seems to take a while as the camera 
initializes.

So what I want to do is as much processing as possible on the existing frame 
before switching to the next frame.  I have an algorithm that can be cut 
short and still be useful.  Rather than alter the parameters to get the frame 
rate I want, I have the frame rate and want to tune the parameters in 
real-time to get as much done as possible without exceeding a 100 msec 
curfew.

I ran into a few strange things when it came to keeping track of the time in 
small increments.  Was hoping someone could explain.

I think the timer call I used was gettimeofday().  One call takes about 2 
usec, which is small compared to the 1300 usec or so for one frame of video 
capture, or the curfew of 100 msec = 100000 usec.

If I make a loop of repeated calls for the time, however, the duration of 
these calls increases.  I made a counter for how many times I could call for 
the time before breaking 100 msec, along with capturing video at that rate.  
It looks like this:

Frame   Loops
0           1
1           399935
2           383604
3           367687
...
21         1178
22          1
23          1

So if you do nothing but ask for the time, like an annoying kid screaming "are 
we there yet?", the response goes from immediate to quite slow.  I don't 
think this happens when there is real processing that puts some delay between 
the calls.  I tried to simulate this with nanosleep() but that actually 
sleeps a good bit more than I requested, so I didn't get very far with a 
control to compare to.

My guess was the kernel throttled back on my process because it was making too 
many calls that required a kernel response.  Not knowing crap about kernels I 
thought I'd raise the question here and see if I could get help.  Hardware is 
Logitech QuickCam 4000, kernel tested was a while back, 2.4.20 or so.

I recently heard that gettimeofday() is only accurate to 18 msec or so.  I 
thought it was working better for me but I need to double-check.  The 
behavior above is still relevant.

Cheers
Sam