predicting leap seconds (was Re: [LEAPSECS] Where the responsibility lies)
On Wed, Jan 04, 2006 at 07:36:17AM +0100, Poul-Henning Kamp wrote: > In message <[EMAIL PROTECTED]>, Neal McBurnett writes: > >On Tue, Jan 03, 2006 at 08:32:08PM +0100, Poul-Henning Kamp wrote: > >> If we can increase the tolerance to 10sec, IERS can give us the > >> leapseconds with 20 years notice and only the minority of computers > >> that survive longer than that would need to update the factory > >> installed table of leapseconds. > > > >Do you have any evidence for this assertion? > > It is an educated guess. > > The IERS have already indicated that they belive they could do prediction > under the 0.9 second tolerance with two or three year horizon. The Torino Colloquium had some discussion of this. Proceedings of the Colloquium on the UTC Timescale held by ITU-R SRG 7A http://www.ien.it/luc/cesio/itu/ITU.shtml Prediction of Universal Time and LOD Variation D. Gambis and C. Bizouard, (IERS) http://www.ien.it/luc/cesio/itu/gambis.pdf After a bunch of nice graphs (not all of which were easy to interpret) I found the periodogram (essentially a discrete Fourier transform of the input data) interesting. The way I read it (expert advice welcomed), the broad peaks at 26 years (0.6 ms/d) and 52 years (0.3 ms/d) suggest that the most common pattern is a gradual cycle a few decades long of lengthening and shortening of the day, presumably driven by movements in the earth's mantle and core. Page 14 of the pdf has a table: Skill of the UT1 prediction statistics over 1963-2003 Horizon Prediction accuracy in ms 3 years 308 2 years 163 1 year 68 180 days 36 90 days21 30 days 7 10 days 3 Perhaps these are worst cases? It would be nice to have confidence intervals. They presented these conclusions: Possibility to predict UT1 with a 1s accuracy at least over 4 years using a simple method : seasonal, bias and drift. New prediction methods are under investigation (Singular Spectrum Analysis, neural network,..) Possibility to use Core Angular Momentum prediction for decadal modeling Steve Allen wrote: > http://www.ucolick.org/~sla/leapsecs/McCarthy.html > > This deserves discussion and analysis and explanation. I wrote Dennis McCarthy about that, and he said he'd look up the details and get back to me next week. But he did remind me of this, which I remember seeing in data they published via ftp years ago: > Regarding the accuracy of these long-term predictions, the IERS > Rapid Service and Prediction Center located at the U. S. Naval > Observatory does make predictions of Delta-T in the IERS Annual > Report. The algorithm for those predictions was determined > empirically by testing a wide range of possibilities. It is > essentially an auto-regressive model using the past ten years of > data. The accuracy based on comparison of observations with what > would have been predicted using that model is shown in the table > below. Note that the accuracy estimates are 1-sigma estimates and > that excursions of 2-sigma (or more) may not be unexpected. > > +-+ > |Year in the Future|Accuracy (1s) (seconds| > |--+--| > |1 | .04 | > |--+--| > |2 | .08 | > |--+--| > |3 | .3 | > |--+--| > |4 | .8 | > |--+--| > |5 | 1. | > |--+--| > |6 | 2. | > |--+--| > |7 | 2. | > |--+--| > |8 | 3. | > |--+--| > |9 | 4. | > +-+ The http://www.iers.org/ points eventually to http://141.74.1.36/MainDisp.csl?pid=47-25786 but the links from there to the annual reports seem broken right now. I still haven't seen any good data on predictions for periods of longer than 9 years. Neal McBurnett http://bcn.boulder.co.us/~neal/
Re: Where the responsibility lies
On Wed 2006-01-04T07:36:17 +0100, Poul-Henning Kamp hath writ: > In message <[EMAIL PROTECTED]>, Neal McBurnett writes: > >Do you have any evidence for this assertion? > > It is an educated guess. > > The IERS have already indicated that they belive they could do prediction > under the 0.9 second tolerance with two or three year horizon. > > >Anyone have a prediction algorithm in mind, and a result of running it > >on the last several decades or centuries of data? > > Makes a great subject for science, doesn't it ? Yes, it does, but the scientist who did the calculation has only barely explained the meaning of the work. http://www.ucolick.org/~sla/leapsecs/McCarthy.html This deserves discussion and analysis and explanation. -- Steve Allen <[EMAIL PROTECTED]>WGS-84 (GPS) UCO/Lick ObservatoryNatural Sciences II, Room 165Lat +36.99858 University of CaliforniaVoice: +1 831 459 3046 Lng -122.06014 Santa Cruz, CA 95064http://www.ucolick.org/~sla/ Hgt +250 m
Re: Where the responsibility lies
In message <[EMAIL PROTECTED]>, Neal McBurnett writes: >On Tue, Jan 03, 2006 at 08:32:08PM +0100, Poul-Henning Kamp wrote: >> If we can increase the tolerance to 10sec, IERS can give us the >> leapseconds with 20 years notice and only the minority of computers >> that survive longer than that would need to update the factory >> installed table of leapseconds. > >Do you have any evidence for this assertion? It is an educated guess. The IERS have already indicated that they belive they could do prediction under the 0.9 second tolerance with two or three year horizon. >Anyone have a prediction algorithm in mind, and a result of running it >on the last several decades or centuries of data? Makes a great subject for science, doesn't it ? Poul-Henning -- Poul-Henning Kamp | UNIX since Zilog Zeus 3.20 [EMAIL PROTECTED] | TCP/IP since RFC 956 FreeBSD committer | BSD since 4.3-tahoe Never attribute to malice what can adequately be explained by incompetence.
Re: Longer leap second notice, was: Where the responsibility lies
> I continue to find the focus on general purpose computing > infrastructure to be unpersuasive. If we can convince hardware and > software vendors to pay enough attention to timing requirements to > implement such a strategy, we can convince them to implement a more > complete time handling infrastructure. This seems like the real goal > - one worthy of a concerted effort. Instead of trying to escape from > the entanglements of this particular system requirement, why don't we > focus on satisfying it in a forthright fashion? As someone who has fought the battles, I can tell you that a simple table is 10x or 100x easier to implement than dealing with parsing the data from N streams. Sure, it limits the lifetime of the device, but a 20 year limit is very reasonable. I had one system that worked over the leap second correctly, even though the code to parse the data from this specific brand of GPS receiver hadn't been written yet. It worked because it knew about the leap second in a table that we'd included on our flash as a fallback when we didn't know anything else. If we could have a table for the next 20 years, there'd be no need to even write the code to get from the GPS stream :-). I know you aren't pursuaded by such arguements. I find your dismissive attitude towards software professionals that have implemented a complete leap second handling infrastructure, with pluggable sources for leap second rather annoying :-( Warner
Re: Where the responsibility lies
On Tue, Jan 03, 2006 at 08:32:08PM +0100, Poul-Henning Kamp wrote: > If we can increase the tolerance to 10sec, IERS can give us the > leapseconds with 20 years notice and only the minority of computers > that survive longer than that would need to update the factory > installed table of leapseconds. Do you have any evidence for this assertion? It seems to me that if IERS had presented a table in 1980, based on the conventional wisdom that the earth is continuing to slow down over time, we'd have been at the edge of that 10-second window in 2000. And who knows how far off a 1995 prediction would be in 2015, or what decadal fluctuations have in store for us in the future. http://hpiers.obspm.fr/eop-pc/earthor/utc/leapsecond.html Anyone have a prediction algorithm in mind, and a result of running it on the last several decades or centuries of data? Neal McBurnett http://bcn.boulder.co.us/~neal/
Re: Longer leap second notice, was: Where the responsibility lies
On Jan 3, 2006, at 4:22 PM, Poul-Henning Kamp wrote: In message <[EMAIL PROTECTED]>, Ed Davies writes: Poul-Henning Kamp wrote: If we can increase the tolerance to 10sec, IERS can give us the leapseconds with 20 years notice and only the minority of computers that survive longer than that would need to update the factory installed table of leapseconds. PHK can reply for himself here but, for the record, I think RS's reading of what he said is different from mine. My assumption is that PHK is discussing the idea that leaps should be scheduled many years in advance. They should continue to be single second leaps - just many more would be in the schedule pipeline at any given point. Obviously, the leap seconds would be scheduled on the best available estimates but as we don't know the future rotation of the Earth this would necessarily increase the tolerance. In theory DUT1 would be unbounded (as it sort of is already) but PHK is assuming that there'd be some practical likely upper bound such as 10 seconds. Am I right in this reading? yes. I'm willing to entertain any suggestion that preserves mean solar time as the basis of civil time. One could view this notion as a specific scheduling algorithm for leap seconds. My own ancient proposal (http://iraf.noao.edu/~seaman/leap) was for a tweak to the current algorithm that would minimize the excursions between UTC and UT1. This suggestion is more than a tweak, of course, since it would require increasing the 0.9s limit. One could imagine variations, however, with sliding predictive windows to balance the maximum excursion against the look ahead time. One is skeptical of any advantage to be realized over the current simple leap second policy. I continue to find the focus on general purpose computing infrastructure to be unpersuasive. If we can convince hardware and software vendors to pay enough attention to timing requirements to implement such a strategy, we can convince them to implement a more complete time handling infrastructure. This seems like the real goal - one worthy of a concerted effort. Instead of trying to escape from the entanglements of this particular system requirement, why don't we focus on satisfying it in a forthright fashion? There is also the - slight - issue that we aren't only worried about "computers". There is a heck of a lot of interesting infrastructure that should be included in the decision making envelope. In general, the strategy you describe could also be addressed as an elaboration on the waveform we are attempting to model with our clocks. Not a constant cadence like tick-tick-tick-tick, that is, but tick-tick-tock-tick. I do think there might be some interesting hay to be made by generalizing our definition of a clock to include quasi-periodic phenomena more complicated than a once-per-second delta function. Would give us some reason to explore the Fourier domain if nothing else. Rob Seaman National Optical Astronomy Observatory
Re: Longer leap second notice, was: Where the responsibility lies
In message <[EMAIL PROTECTED]>, Ed Davies writes: >Poul-Henning Kamp wrote: >>> If we can increase the tolerance to 10sec, IERS can give us the >>> leapseconds with 20 years notice and only the minority of computers >>> that survive longer than that would need to update the factory >>> installed table of leapseconds. > >PHK can reply for himself here but, for the record, I think RS's >reading of what he said is different from mine. My assumption is >that PHK is discussing the idea that leaps should be scheduled many >years in advance. They should continue to be single second leaps - >just many more would be in the schedule pipeline at any given >point. > >Obviously, the leap seconds would be scheduled on the best available >estimates but as we don't know the future rotation of the Earth this >would necessarily increase the tolerance. In theory DUT1 would be >unbounded (as it sort of is already) but PHK is assuming that there'd >be some practical likely upper bound such as 10 seconds. > >Am I right in this reading? yes. -- Poul-Henning Kamp | UNIX since Zilog Zeus 3.20 [EMAIL PROTECTED] | TCP/IP since RFC 956 FreeBSD committer | BSD since 4.3-tahoe Never attribute to malice what can adequately be explained by incompetence.
Re: Longer leap second notice, was: Where the responsibility lies
Poul-Henning Kamp wrote: If we can increase the tolerance to 10sec, IERS can give us the leapseconds with 20 years notice and only the minority of computers that survive longer than that would need to update the factory installed table of leapseconds. Rob Seaman replied: No. Rather all computers that exist during such an event are obligated to deal with it. The number of deployed systems follows some increasing trend similar to Moore's law. By delaying the adjustments, you guarantee that more systems will be affected when they do occur. And, unless you can guarantee that a particular deployed system (and systems derived through various upgrade pathways) will be retired prior to the adopted horizon, prudent policy would require remediation in any event. Would like to see a proposed architecture a little more detailed than a "factory installed table". PHK can reply for himself here but, for the record, I think RS's reading of what he said is different from mine. My assumption is that PHK is discussing the idea that leaps should be scheduled many years in advance. They should continue to be single second leaps - just many more would be in the schedule pipeline at any given point. Obviously, the leap seconds would be scheduled on the best available estimates but as we don't know the future rotation of the Earth this would necessarily increase the tolerance. In theory DUT1 would be unbounded (as it sort of is already) but PHK is assuming that there'd be some practical likely upper bound such as 10 seconds. Am I right in this reading? Ed.
Re: Where the responsibility lies
All right - I guess we can go another round or two while waiting - perhaps indefinitely - for reports of leap second related catastrophes to filter in. First, an apology for posting my previous reply publicly. It escaped my notice that I was replying to a private message. On Jan 3, 2006, at 12:32 PM, Poul-Henning Kamp wrote: If you already have to cope with DUT1 anyway, how much difference can it possibly make if the definition says |DUT1| < .9 or |DUT1| < 10sec or |DUT1| < 1 hour The amplitude of the effect is 3600 times as great. The interval between each event much longer (though quadratic, not proportionate, of course). The larger amplitude implies a larger impact on human activities. The longer interval implies that future instrumentalities will be less prepared to deal with the impact. If we can increase the tolerance to 10sec, IERS can give us the leapseconds with 20 years notice and only the minority of computers that survive longer than that would need to update the factory installed table of leapseconds. No. Rather all computers that exist during such an event are obligated to deal with it. The number of deployed systems follows some increasing trend similar to Moore's law. By delaying the adjustments, you guarantee that more systems will be affected when they do occur. And, unless you can guarantee that a particular deployed system (and systems derived through various upgrade pathways) will be retired prior to the adopted horizon, prudent policy would require remediation in any event. Would like to see a proposed architecture a little more detailed than a "factory installed table". As far as I know, less than 1% of people on this planet actually have the sun straight south at 12:00 local time today and a quite sizeable minority (a lot of China) lives perfectly happy with the sun being further than 15 degrees from south at local 12:00. Once again confusing secular and periodic effects. Leap seconds are the result of a monotonic trend. Time zones and daylight saving repeat seasonally or represent geographic constants. It follows from this, that a proposal for a 1hour tolerance on DUT1 is perfectly feasible without odd things happening to the cows milk etc. Got me. Might it not be prudent to ask a farmer? We interviewed a guy last year who networks dairy cattle via instrumented capsules that monitor pH and communicate wirelessly from cow to cow back to the barn. Farmers are likely among the more technologically literate community for temporal data products. Why not benefit from their expertise? I think the leap hours is a political tool to make the proposal go through without commiting anybody to anything for the next couple of hundred of years. Of course. We differ on whether it is: A) ethical, or B) viable. There are three orders of magnitude difference between a leap second and a leap hour, and consequently the need for leap hours will grow less rapidly than the need for leapseconds. Bzzz! Nice try, but an incorrect answer. Thanks for playing our home game. I've appended Steve Allen's excellent plot of the long term behavior of length of day. LOD grows linearly over scales pertinent to this discussion (discounting wiggles due to interesting geophysics). This is due, of course, to angular momentum transfer from the Earth's rotation to the Moon's orbit. Mitigating factors include the rebound of the continents since the last ice age. Leap seconds accumulate as the integral under this curve (thus quadratically). The last hour's worth took 1200 years. (Hence the 600 year estimate for issuing the first leap hour to limit excursions to +/- 0.5 hours.) The hour before that, only four centuries. It does not matter whether these are binned as leap seconds or leap hours, they will arrive when they arrive. That being the case, one argues that bleeding off the accumulation in smaller, more frequent doses is a better choice. Rob Seaman National Optical Astronomy Observatory --- ancient.pdf Description: Adobe PDF document
Re: Where the responsibility lies
In message <[EMAIL PROTECTED]>, Rob Seaman writes: >John Hawkinson replies: > >I think PHK has demonstrated the ability (and willingness :-) to hold >up his own end of an argument. Should we ever find ourselves at the >same conference, I'll buy him a beer in anticipation of a rousing >discussion. I'll be happy to bring booze as well :-) >There are several issues confounded here. First, an untested >assertion that eliminating leap seconds will simplify time handling. >DUT1 looms large in astronomical software and one would have to be >convinced that this is not an issue with other disciplines. But it's exactly the fact that DUT1 already exists that bugs me. If you already have to cope with DUT1 anyway, how much difference can it possibly make if the definition says |DUT1| < .9 or |DUT1| < 10sec or |DUT1| < 1 hour If we can increase the tolerance to 10sec, IERS can give us the leapseconds with 20 years notice and only the minority of computers that survive longer than that would need to update the factory installed table of leapseconds. >Third, leap seconds are a mechanism to realize mean solar time in >practice. As would leap-hours (or jumping timezones) be. It's only a matter of the tolerance we accept on DUT1. As far as I know, less than 1% of people on this planet actually have the sun straight south at 12:00 local time today and a quite sizeable minority (a lot of China) lives perfectly happy with the sun being further than 15 degrees from south at local 12:00. It follows from this, that a proposal for a 1hour tolerance on DUT1 is perfectly feasible without odd things happening to the cows milk etc. >The acknowledgment of a contingent need for >leap hours shows that the authors of the ITU proposal understand this. I think the leap hours is a political tool to make the proposal go through without commiting anybody to anything for the next couple of hundred of years. >Fourth, the need for leap seconds is growing quadratically as the >Earth continues to slow. We have no business making ad hoc policies >based on the rarity of events that are becoming more frequent. The >need for "leap hours" will grow just as rapidly - and much more >dramatically. A solution that ignores real world constraints is no >solution. Uhm, no. There are three orders of magnitude difference between a leap second and a leap hour, and consequently the need for leap hours will grow less rapidly than the need for leapseconds. >In short, leap hours are - well - dumb. A proposal that relies on >their use, naive. Leap hours or leap seconds is only a matter of magnitude and frequency and consequently both are equally naïve. Poul-Henning -- Poul-Henning Kamp | UNIX since Zilog Zeus 3.20 [EMAIL PROTECTED] | TCP/IP since RFC 956 FreeBSD committer | BSD since 4.3-tahoe Never attribute to malice what can adequately be explained by incompetence.
Re: Where the responsibility lies
John Hawkinson replies: Time handling bugs typically appear in the interfaces between systems that make contradictory assumptions. I think phk's point ("text book example") was that these problems were more likely to have been detected in a world where everyone's default timescale (UTC) was not subject to leap seconds -- rare events that may not come up in the course of normal testing unless special care is given. I think PHK has demonstrated the ability (and willingness :-) to hold up his own end of an argument. Should we ever find ourselves at the same conference, I'll buy him a beer in anticipation of a rousing discussion. There are several issues confounded here. First, an untested assertion that eliminating leap seconds will simplify time handling. DUT1 looms large in astronomical software and one would have to be convinced that this is not an issue with other disciplines. Second, that UTC is indeed "everyone's default". It is (rather loosely) the current civil time standard, but I fail to see why this makes it a default choice for a precision timing application. "Civil measurement" in the USA still revolves around twelve inches to a foot, but SI units are the default for American scientists as is true elsewhere in the world. I won't belabor my assertion that mean solar time is the obvious civil time standard. But why do we expect that an X-ray spacecraft would of necessity adopt such a standard? Third, leap seconds are a mechanism to realize mean solar time in practice. The underlying issues are those of solar time, not how this is achieved. One can assert that leap seconds should cease, but the fundamental civil timing requirements will remain in force. Some other mechanism for synchronizing our clocks to our home planet must needs be identified. The acknowledgment of a contingent need for leap hours shows that the authors of the ITU proposal understand this. Fourth, the need for leap seconds is growing quadratically as the Earth continues to slow. We have no business making ad hoc policies based on the rarity of events that are becoming more frequent. The need for "leap hours" will grow just as rapidly - and much more dramatically. A solution that ignores real world constraints is no solution. Fifth, normal testing should involve special care - or what is its value? In short, leap hours are - well - dumb. A proposal that relies on their use, naive. Rob Seaman National Optical Astronomy Observatory
Where the responsibility lies (was Re: [LEAPSECS] text book example...)
On Jan 1, 2006, at 3:29 AM, Poul-Henning Kamp wrote: http://lheawww.gsfc.nasa.gov/users/ebisawa/ASCAATTITUDE/ This describes a system for "attitude determination", i.e., for pointing the ASCA X-ray telescope at celestial objects. It appears there were several bugs in time handling. They get full marks for diagnosing, documenting, and publicizing the resulting issues. These include issues of data characterization as well as of spacecraft operations. As someone already pointed out, the fundamental issue was the adoption of UTC as a standard when TAI, for instance, might have been more appropriate. UTC conveys Earth orientation information - one might wonder why it would be the obvious choice for a platform not located on Earth. The blind adoption of any standard is a bad thing. Adopting American standards instead of metric, or worse yet, adopting both, can be fatal to a project. Give them an inch and they'll take a metre. Posit for the moment that this project had completely avoided the use of UTC. Since the mission required maintaining attitude with respect to the celestial sphere and solar system objects, requirements for handling DUT1 (or the equivalent) would have remained. Having implemented an astrometric software package (others here are much more experienced), I can assure you that such software is no less subject to the creation of obscure bugs than any other system. The bugs that were realized as reflections of poor leap second handling might well have instead appeared in other ways. UTC is a handy tool for conveying both UT1 and TAI in one concise package. Depending on the project, various assumptions and approximations may be appropriate. It sounds like inappropriate assumptions were made. Oh well. Time handling bugs typically appear in the interfaces between systems that make contradictory assumptions. The quoted page's description of two bugs that originally cancelled out later reasserting themselves with a maintenance update is classic for such things. Fixing one bug reveals others that were previously masked. Bugs don't result from standards, they are the responsibility of engineers. The existence of a bug certainly isn't an argument for ignoring underlying facts of nature. The ASCA mission ended in 2000 as the result of a geomagnetic storm (another pesky fact of nature keyed to Earth orientation) and the spacecraft burned up on reentry in 2001. (Earth orientation is certainly a matter of public interest as to where de-orbiting spacecraft might impact.) However, we have yet to hear of any significant issues encountered during the 2005 leap second by current (appropriate or inappropriate) users of UTC. Rob Seaman National Optical Astronomy Observatory