[LEAPSECS] When did computer timekeeping get good enough for leap seconds to matter?

[Warner Losh]

On Jan 9, 2014, at 4:03 AM, Hal Murray wrote:


> 

> The IBM 360 systems starting in 1964 used the power line frequency.  (A 

> location in low memory got bumped at 300 counts per second.  5 per cycle on 

> 60 Hz and 6 per cycle on 50 Hz.)  I wonder how much the power timekeeping 

> wandered back then relative to today.

> 

> Does anybody know what the guys in the power company control rooms do about 

> leap seconds?

> 

> ------------

> 

> Leap seconds started in 1972.

> 

> I was at Xerox in the late 1970s.  At boot time, Altos got the time from a 

> local time server.  Altos used the system crystal (5.88 MHz) for timekeeping. 

> Personal Altos were rebooted frequently so it didn't matter if their clock 

> drifted a bit.  The time server was packaged with the routers.  (We called 

> them gateways.)  On the few systems that were up a long time (file servers, 

> routers), we hand tweaked a fudge factor to adjust the clock rate.  It wasn't 

> hard to get to a second per week.  I think the units for the fudge factor 

> (from a config file) were seconds per day, but it would read at least one 

> digit past the decimal point.  I don't remember any mention of leap seconds.

> 

> 

> When were there enough (Unix?) boxes on the net running NTP and keeping good 

> enough time to notice things like leap seconds?

> 

> I should go browse the old RFCs and see when the API for telling the kernel 

> about pending leap seconds was published.  But somebody may have good stories 

> or folklore.


I know there were documented problems in the leap seconds that happened in the late 1990s. I was involved in GPS steered OCO in the early 2000's, and they were definitely a problem by then. That's when i developed most of my opinions about their impact on general time keeping and imperfect fit with POSIX and the leap second standard. A fit that's only grown more chafing to this day.

Warner