2-3 is fine.but the problem is when it's actually put to the test in userspace where you have many threads competing for cpu, that is where you have to measure the accuracy which is what that test during startup is doing. using FS to sleep 1000 microseconds and timing how close it actually is to that.
voip can withstand jitter on most implementation as long as the running average of the packets is on the mark, the timing of each individual packet is less important, this worsens over latent connections.
try this from the cli
time_test 1000
you will see it come out + or - 100 easily (there is overhead for the code executing as well as other factors that can skew the results)
then try it with a larger sample:
time_test 1000 1000
time_test 1000 2000
etc
the more times you try it the more likely the average in the long run will be accurate which is all we can hope for in our pseudo real time environment.
most times if you try this with -nocal you will always get 2000 when you intended for 1000
which means all your 1ms sleeps are double length from what was requested.
some newer kernels try to make it more accurate but by doing so they eat up so much cpu it's not worth it. For the most part semi-accurate 1ms timing is more than enough.