I connected pin 10(ACK) to pin 3 (data out). Then ran 2 copies of 
testpi. testpi #1 used to "get.time", and the other generated pulses ("pulse").

I then graphed the ISR_timestamp - pulse_timestamp.

Ran these test on a CentOS 5/2.6.18-194.17.1.el5 kernel with a 
AMD Sempron(tm) Processor 3300+ CPU on 28Oct2010.

1. rep=1000 usec=1000000  (1.0 sec x 1000)

   Graph of ISR_timestamp - pulse_timestamp shows worst case of 10 usec.
    
2. rep=10000 usec=100000 (0.100 sec x 10000) Graph of ISR_timestamp - pulse_timestamp shows worst case of 11 usec.
3. rep=100000 usec=10000 (0.010 sec x 100000) Graph of ISR_timestamp - pulse_timestamp shows worst case of 14 usec.
4. rep=100000 usec=1000 (0.001 sec x 100000) (The pulses loop actually ran at about 0.002 ). Graph of ISR_timestamp - pulse_timestamp shows worst case of 300 usec. All other samples are under 40 usec.
"Looks like parint could provde ~100 resolution microseconds 99.99% of the time."