40ms sounds okay for me, even for most command line tools. Oh and I most probably don't get a segmentation fault ;-)

    $ time ruby -e "puts 'hello world'"
    > hello world
    >
    > real	0m0.076s
    > user	0m0.029s
    > sys	0m0.017s

    $ time ruby -e "puts 'hello world'"
    > hello world
    >
    > real	0m1.161s
    > user	0m2.050s
    > sys	0m0.093s

Here's executing an assembly (über-jar) I wrote in Scala:

  → time scripts/couch 
  Please provide a basePath and a valid command. ("backup", "restore", "truncate", "migrate" or "migrate-all")

  real	0m0.375s
  user	0m0.415s
  sys	0m0.073s

A few hundred ms on my 1.2GHz MacBook. And this includes a bunch of Akka Actors.

E.g. doing a rolling update across a large number of instances by restarting a service at a time can become a quick enough process to be viable in instances where you'd otherwise need lots of excess capacity to be able to cycle larger proportions of instances at the same time. Making full rolling updates "cheaper" both in time and resources also translates to making rapid updates a safer choice (e.g. if I can roll back a broken release in 5 minutes, it's far safer to push out a new release than if a rollback takes hours).

There is no problem with restarting nginx to update its config once a second, and it's even encouraged that you do so. Besides, crash-only software recovery is really the only morally sound development technique.