Rapidly issue a large number of move orders to different overlapping subsets of those units, each with a different destination. (Actually, this should work just as well if you rapidly re-issue the same move order to a large subset of your units.)
Observe that the server hangs for seconds or minutes trying to do all the pathfinding.
This demonstrates a few related problems:
We really need a unit cap.
When a unit is ordered and then re-ordered in rapid succession, we do the complete pathfinding for the first destination and then again for any subsequent destinations, even if the unit doesn't actually get a chance to do any moving toward the first one.
When a large number of units in the same chunk are ordered to the same other chunk, we do the same chunk-to-chunk pathfinding calculation repeatedly for all of those units. We should be able to memoize it better.
Try the following:
This demonstrates a few related problems: