Harlan, Why should your head explode? Possibly you’re overthinking the problem. And there is no reason (or simple way I can envision) to test my plan, as you advise, in advance. I will just block NTP in my border router temporarily. No need to make a mountain out of this molehill. Cisco, and many other NTP client gear vendors, recommend this approach, and they’ve published extensive research on the matter.
-mel > On Jun 23, 2015, at 12:46 AM, Harlan Stenn <st...@ntp.org> wrote: > > This stuff can make my head explode. > > When a leap second is added, like on 30 June 2015 at the last second of > the day, POSIX insists that the day still have 86400 seconds in it. > This makes the day longer by one second, so time has to either slow down > or move backwards. > > The "dumb" way to do this is to step the clock back by 1 second at the > instant before the stroke of midnight. > > The allegedly better way to do this would be to stop the clock a bit > before midnight, and hold the time for 1 second. To continue providing > monotonic time, every time somebody says "what time is it" during that > holding period one would want to bump the time by the smallest amount > possible, usually 1 nanosecond (assuming the kernel keeps time in > nanoseconds). > > Ideally you wouldn't want to add enough nanoseconds to cause the clock > to roll over into the next day "too early". > > But apparently nobody has implemented this, even though Prof. Mills > described it in RFC-i-forget about 20 years ago. > > This is mostly because POSIX deals with absolute time and not relative > time. > > In the unlikely event of a leap second deletion, there would be no > 23:59:59, so when the clock is about to strike 23:59:59 it's OK to add > an extra second to the clock to effectively have the time "jump" from > 23:59:58 to 00:00:00. This is still a monotonic increment in time. > > Whatever you decide to do, I recommend you actually test it out to see > if it behaves the way you think it will. You have a whole week still. > > H
