FYI. Excerpts from RISKS DIGEST 17.08, 17.09, 17.10, 17.11 ----------------------------- Date: Sun, 23 Apr 1995 22:02:42 +0059 (EDT) From: Robert J Horn <rjh@world.std.com> Subject: Floating-Point Time The opponents of floating-point representation for time have done an insufficient analysis. About twenty years ago I was part of a research group doing extensive time series analysis of weather and related data. We needed a good way to represent time. Fortunately we had a few astronomers on the team, so time was reasonably well understood. We chose "second of century", using a double precision floating point representation. Analysis showed that this would preserve millisecond accuracy for the span of interest. (Actually for all of recorded history and more.) Since we usually were satisfied with one minute accuracy this seemed sufficient. There was a brief debate about using a better time base, but 12:00:01 AM GMT, 1 January, 1901 was easy to explain to everyone. There are a few applications that need better than millisecond precision, but for most of the worlds applications double precision floating point will provide enough precision for the next few millenia. (A simple test for those who are unsure about their needs. Do you compensate for the variations in the rate of the Earth's rotation? If not, you probably don't need millisecond accuracy.) This notation had some interesting side effects. At the time, floating point turned out to be somewhat faster than 64-bit integers due to a quirk of hardware. It also led to excellent compatibility with the other time series processing. Time was just another well behaved variable. This notation eliminated a lot of the mistakes made by the typical programmer who is ignorant of traditional time notations and their problems. There could have been some round-off issues, but we rarely did any arithmetic other than addition or subtraction of two times, where millisecond accuracy is maintained. It even led to a simple notation for interval time span data, e.g. "0.01 inches of rain fell between 1633 and 1647 on ...", which is how many meteorological measurements are made. The difficult problems were in translation to and from local. The most severe problem was the inherent ambiguity of local time in recent decades. There are two true times corresponding to each time in the one hour of overlap when Daylight Savings shifts back to Standard. Correctly resolving this ambiguity was always a headache. Fortunately most professional measurements have been recorded in UTC, or GMT before UTC was defined. A word of caution, double precision floating point is suitable for an internal representation of UTC, or "absolute" time. You have to do your own analysis if you are interested in timing relative to some event. Rob Horn rjh@world.std.com P.S. The turn of century problem has made The NY Times. It may be so widely hyped that almost all the problems are fixed by the time it comes. [Hmm! According to you, it comes at 1/1/01 rather than 1/1/00. I wonder who agrees with that! PGN] ----------------------------- ----------------------------- Date: Tue, 25 Apr 1995 10:54:01 -0400 From: rjh@world.std.com (Robert J Horn) Subject: Re: Floating-Point Time In comp.risks you write:
[Hmm! According to you, it comes at 1/1/01 rather than 1/1/00. I wonder who agrees with that! PGN]
Who am I to argue with astronomers? They work in mathematics and FORTRAN, so counting starts with one. Therefore the first day of the first century is 1/1/00000001. C was invented later, and never really accepted by astronomers. We just moved things closer by starting at 1901. Rob Horn rjh@world.std.com [The customary convention was also commented on by jcs@zoo.bt.co.uk (John Sager), pbm1001@thor.cam.ac.uk (Paul Menage), harper@kauri.vuw.ac.nz (John Harper), ag325@detroit.freenet.org (William M. Bickham), "david (d.p.) woodman" <woodman@bnr.ca>, rjwells@undergrad.math.uwaterloo.ca (justin wells), and Greg Lindahl <gl8f@fermi.clas.virginia.edu>, an astronomer who noted "Astronomers get to go to 2 sets of turn-of-the-century parties... you nay-sayers only get to go to one." I'll be at the former, when most of the computer-related risks are likely to begin. PGN] ----------------------------- Date: Tue, 25 Apr 95 21:02:17 PDT From: "Peter G. Neumann" <neumann@csl.sri.com> Subject: Re: Floating-Point Time Oh, yes, I certainly agree that the astronomers and ephemeris/almanac folks like 1/1/01 as the century start. However, because there was no year ZERO, that does not scale backwards. The first century BC clearly began on 1/1/-100, and the first millennium BC on 1/1/-1000. The only SANE way to handle this is to provide a 99-year first century; just as we have leap-years and leap-seconds, we could have a backwards leap-century. Indeed, it is just as well there were no computers in virtual-year 0000, or the religious wars would have been proven recursively unsolvable. PGN ----------------------------- Date: Tue, 25 Apr 95 13:53:39 PDT From: geoff@ficus.CS.UCLA.EDU (Geoff Kuenning) Subject: Re: Floating-Point Time (Robert J Horn, RISKS-17.08)
We chose "second of century", using a double precision floating point representation. Analysis showed that this would preserve millisecond accuracy for the span of interest.
Sigh. There is more to time calculations than just understanding time, and there is more to numerical analysis than just the range of the mantissa. I'll try this once more, keeping it brief. This project wouldn't have used IEEE format 20 years ago, but let's proceed with the analysis under modern assumptions. The IEEE double-precision floating-point representation provides 53 bits of significance in the mantissa. Using the approximation that 2^10 = 10^3, this can be see to allow a range of about 8x10^15 in the mantissa, before bits start getting dropped. Since there are about 3x10^7 seconds in a year, or about 10^8 every 3 years, one can represent about 8x16x3 = 384 years to millisecond precision without violating that range, right? Wrong. This is true only if you represent time as integer milliseconds. Since the representation used seconds, the milliseconds were represented fractionally. There are only four millisecond values that can be represented accurately in a binary floating-point system: 0, 250, 500, and 750. ANYTHING ELSE is an approximation. This is especially true if your calculations involve any addition and subtraction of times.
Since we usually were satisfied with one minute accuracy this seemed sufficient.
It sounds like this project involved some sort of modeling of the physical world. In such cases, the loss of a few bits of accuracy may not matter, although I still think programmers ought to understand numerical analysis better than most do. But remember that the original discussion of time representation arose in the context of measuring time in computer and financial applications, where these bits can matter.
There are a few applications that need better than millisecond precision, but for most of the worlds applications double precision floating point will provide enough precision for the next few millennia. (A simple test for those who are unsure about their needs. Do you compensate for the variations in the rate of the Earth's rotation? If not, you probably don't need millisecond accuracy.)
This is both a broad and a biased statement. Here's a simple response to the simple test: I'm not dealing with astronomical phenomena, and so I couldn't care less about variations in the Earth's rotation. What I care about is whether I can query the system time, do something, query the time again, and subtract to get an accurate elapsed time. I care about not only milliseconds, but microseconds (and in a few years, nanoseconds). Now it's true that I generally deal with time in a small range. Unfortunately, the system time is represented in terms of a relatively long base interval. So I have to be able to take two times measured in years, subtract them, and get microsecond precision in my results. If the times are represented as integer microseconds, I can be sure that everything will "balance," as the accountants say. If a sloppy floating-point representation insists on giving it to me as integer and fractional seconds, things won't add up, and my papers may not get published.
There could have been some round-off issues, but we rarely did any arithmetic other than addition or subtraction of two times, where millisecond accuracy is maintained.
But addition and subtraction are precisely the places where numerical errors are introduced! Millisecond accuracy was *not* maintained in this situation. Since you only cared about minutes, this hardly mattered, but it's still an incorrect statement.
A word of caution, double precision floating point is suitable for an internal representation of UTC, or "absolute" time.
It's only appropriate when you're doing modeling of the physical world, where you don't care about losing a few bits of accuracy in the fraction because your measurements aren't that good anyway. Double precision is a *terrible* way to model elapsed time inside a computer system. It has nothing to do with the nature of time, and everything to do with numerical analysis. Geoff Kuenning g.kuenning@ieee.org geoff@ITcorp.com http://www.cs.ucla.edu/ficus-members/geoff/ ----------------------------- ----------------------------- Date: Sat, 29 Apr 1995 19:49:08 GMT From: dcline@netcom.com (David Cline) Subject: Re: Floating-Point Time (Kuenning, RISKS-17.09)
... Since there are about 3x10^7seconds in a year, or about 10^8 every 3 years, one can represent about 8x16x3 = 384 years to millisecond precision without violating that range, right?
Wrong. This confuses milliseconds and microseconds; You can represent 285 years to *microsecond* accuracy in 53 bits. If you only care about millisecond accuracy, you can represent about 285,000 years. There are also ways of using the sign bit to double the effective range. Dave Cline Spring Valley Software dcline@netcom.com [Your moderator is dismayed that this is dragging on so long! PGN] ----------------------------- Date: Fri, 28 Apr 95 11:12:30 EDT From: hopkins@VFL.Paramax.COM Subject: Re: Floating-Point Time On the year-zero and religious wars: PGN suggests [RISKS-17.09] that first-century dates (which were, after all, not invented until well after the fact) would have created religious wars had there been computers to suggest that there should be a year zero. Any self-respecting computer, however, would have balked at attempts to divide the factions by zero. Bill Hopkins hopkins@VFL.Paramax.Com Unisys Corporation (Soon to be Loral, they say) 610-648-2854 or 363-7464 Valley Forge Eng'g Ctr, POB 517, Paoli PA 19301 ----------------------------- ----------------------------- Date: Tue, 2 May 1995 19:49:51 -0400 From: "Andrew D. Fernandes" <adfernan@cnd.mcgill.ca> Subject: Re: Floating-point time It seems to me that the real issue as to whether or not floating point representations are appropriate for timestamps is not "can I get microsecond resolution?" but "is it going to work unconditionally?" Floating point numbers bedevil numerical programmers for several reasons; we need to know the radix, number of mantissa and exponent digits, behaviour of rounding, etc. to guarantee "good" results---there is a reason that textbooks have been written about numerical analysis. Modern computers generally stick to the IEEE-754 standard for floating point arithmetic, and thus code is usually fairly interchangeable, but anyone who has ever done intensive number crunching on old IBMs, VAXen, or CRAYs can tell horror stories about interchanging floating point programs between systems. Even going between a SPARCstation and my '486 at home, I can see a few digits change. "Unconditional portability" is ludicrous under these conditions. A 64 bit integer can represent 1.84e19 values, which implies about 2 nanosecond resolution of a thousand year interval. Integer math is very portable across computer systems, or at least is very easy to describe fully, and 1 + 1 always evaluates as 2, and not 2.0001. -Andrew D. Fernandes (adfernan@cnd.mcgill.ca) ----------------------------- Date: Thu, 4 May 95 11:42 PDT From: ludemann@netcom.com (Peter Ludemann) Subject: Re: Floating-point time In my original note I mentioned something really simple: handling clock ticks. Not handling time, which is a topic for PhD theses (Stanford is about to grant a PhD for a thesis on temporal representation in databases). My problem was simple: I needed to do 48-bit arithmetic on a machine which had 32-bit integers. I had been given an assembler routine which did the 48-bit arithmetic wrong; and I was fortunate to test it in December when its accumulated error was about 3 minutes in converting clock ticks to a date (in January, it had almost no error). So, what to do? Obviously, writing 48-bit arithmetic in assembler was error-prone and tedious. Doing it in PL/I (this was before C was generally available) wasn't much more fun nor less error-prone (anyone who has read the PL/I manual's description of handling overflow and precisions will agree that it's not a pretty sight; and I later learned that the PL/I implementors had got it wrong anyway). And I didn't have access to LISP's rational-number packages. In a flash of inspiration, I realized that double-precision floating point, with 53-bit mantissas, would handle my 48-bit integers with no round-off errors. Converting between 48-bit integers and floating point is simple (it's a "well-known assembler idiom" that takes about 3 instructions). And everything would then proceed perfectly (even a Pentium can get floating-point addition and subtraction right). At one stroke I had removed a few pages of difficult assembler and replaced them by a few lines of PL/I. Less code, less chance of error. [Inspired by this, I wrote a disk accounting package that eschewed packed decimal arithmetic and instead used floating point (of course, I calculated everything in cents because 1.01 doesn't represent exactly but 101 does. Another success.)] One of my jobs as a programmer is to reduce the risk of error in my system. I can only test so much. Testing 48-bit integer arithmetic is not fun, at least for me; and proving the code right is difficult (Knuth allegedly wrote "Beware of bugs in the above code; I have only proved it correct, not tried it."). PL/I's implementation of 15-digit packed decimal was buggy. And I wasn't being paid by the number of lines of code (au contraire: if I finished early, I took time off). The lessons for RISKS: before you try to do something complicated, maybe there's something simple and reliable lying around that you can pervert slightly into the form you want. [My original RISKS lesson was to be sure to test time/date conversion routines at many times of the year; if they work in January, they might not work in December.] - Peter Ludemann ----------------------------- Date: Tue, 2 May 1995 18:26:06 GMT0BST From: Phil Brady <P.R.Brady@swansea.ac.uk> Subject: Re: Floating-point time [Your moderator is dismayed that this is dragging on so long! PGN] Agreed - the debate has been interesting, but hasn't it run it's course yet? The fact is that every programmer needs to decide whether to use floating, decimal, integer, string, etc and the appropriate precision for every variable in every program depending on the application. If they don't know the appropriate form for time for the problem in hand, then they aren't programmers! Phil Brady, University of Wales, Swansea 01792 295160 ----------------------------- ----------------------------- Date: 24 March 1995 (LAST-MODIFIED) From: RISKS-request@csl.sri.com Subject: Info on RISKS (comp.risks), contributions, subscriptions, FTP, etc. The RISKS Forum is a moderated digest. Its USENET equivalent is comp.risks. Undigestifiers are available throughout the Internet, but not from RISKS. SUBSCRIPTIONS: PLEASE read RISKS as a newsgroup (comp.risks or equivalent) on your system, if possible and convenient for you. BITNET folks may use a LISTSERV (e.g., LISTSERV@UGA): SUBSCRIBE RISKS or UNSUBSCRIBE RISKS. U.S. users on .mil or .gov domains should contact <risks-request@pica.army.mil> (Dennis Rears <drears@pica.army.mil>). UK subscribers please contact <Lindsay.Marshall@newcastle.ac.uk>. Local redistribution services are provided at many other sites as well. Check FIRST with your local system or netnews wizards. If that does not work, THEN please send requests to <risks-request@csl.sri.com> (which is not yet automated). SUBJECT: SUBSCRIBE or UNSUBSCRIBE; text line (UN)SUBscribe RISKS [address to which RISKS is sent] CONTRIBUTIONS: to risks@csl.sri.com, with appropriate, substantive Subject: line, otherwise they may be ignored. Must be relevant, sound, in good taste, objective, cogent, coherent, concise, and nonrepetitious. Diversity is welcome, but not personal attacks. PLEASE DO NOT INCLUDE ENTIRE PREVIOUS MESSAGES in responses to them. Contributions will not be ACKed; the load is too great. **PLEASE** include your name & legitimate Internet FROM: address, especially from .UUCP and .BITNET folks. Anonymized mail is not accepted. ALL CONTRIBUTIONS CONSIDERED AS PERSONAL COMMENTS; USUAL DISCLAIMERS APPLY. Relevant contributions may appear in the RISKS section of regular issues of ACM SIGSOFT's SOFTWARE ENGINEERING NOTES, unless you state otherwise. All other reuses of RISKS material should respect stated copyright notices, and should cite the sources explicitly; as a courtesy, publications using RISKS material should obtain permission from the contributors. RISKS can also be read on the web at URL http://catless.ncl.ac.uk/Risks Individual issues can be accessed using a URL of the form http://catless.ncl.ac.uk/Risks/VL.IS.html (Please report any format errors to Lindsay.Marshall@newcastle.ac.uk) RISKS ARCHIVES: "ftp unix.sri.com<CR>login anonymous<CR>[YourNetAddress]<CR> cd risks<CR> or cwd risks<CR>, depending on your particular FTP. Issue J of volume 17 is in that directory: "get risks-17.J<CR>". For issues of earlier volumes, "get I/risks-I.J<CR>" (where I=1 to 16, J always TWO digits) for Vol I Issue j. Vol I summaries in J=00, in both main directory and I subdirectory; "bye<CR>" I and J are dummy variables here. REMEMBER, Unix is case sensitive; file names are lower-case only. <CR>=CarriageReturn; UNIX.SRI.COM = [128.18.30.66]; FTPs may differ; Unix prompts for username and password. Also ftp bitftp@pucc.Princeton.EDU. WAIS repository exists at server.wais.com [192.216.46.98], with DB=RISK (E-mail info@wais.com for info) or visit the web wais URL http://www.wais.com/ . Management Analytics Searcher Services (1st item) under http://all.net:8080/ also contains RISKS search services, courtesy of Fred Cohen. Use wisely. -----------------------------