Here is an example:
All the chopexecDATs are doing similar stuff - setting constCHOP parameter.
At each container one of chopexec has a twice more CPU cooking time, at some circumstances (seems randomly) it could reach up to 0.04 ms average time.
The reason one is always taking longer is that there is some data propagating to other nodes that only happens once. Whichever node happens to cook first (which is a bit arbitrary) does all the work necessary and then the others process faster.
I understand why this seems like a bug! It confused me at first when I looked at it too.