To understand the parameter knobs, these were the rules of thumbs that 
motivated them....

1.  Set hold-times to match typical round-trip times to allow the "system" 
to respond before acting again (~100ms in the CDN world of today)

2.  Set increments to allow the marking probability to go from 0 to 1 over 
the timescales that aggregate loads typically change (from several seconds 
to minutes).  As an example, if your link can go from zero to saturated on 
the order of 10 seconds, then you want the marking probability to adapt from 
0-1 in 10 seconds.  Assuming a 100ms hold-time, you'd then want to use an 
increment of 0.01 to do this.

3.  Set decrements similarly, but make sure they are much smaller than the 
increments.  The reason this is the case is that packet losses only happen 
when the marking probability is too low while underutilization occurs when 
the marking probability is either too low or too high.