Please don’t make benchmarks with timing inside the loop creating a sum. Just time the loop and divide by the number. Stuff happens getting the time and the jitter can mess with results.
The real world benchmark is measuring it from invocation, both for cold launches and 'hot' (data cached from the last run).
Interestingly I might have only ever used the time (shell) builtin command. GNU's time measuring command prints a bunch of other performance stats as well.
