The "Inaccurate Clock" Mini-HOWTO ================================= By rbean@execpc.com last revised Dec. 1996 The real-time-clock chips used on PC motherboards (and even expensive workstations) are notoriously inaccurate. Linux provides a simple way to correct for this in software, making the clock potentially *very* accurate even without an external time source. But most people don't seem to know about it, for several reasons: 1: It's not mentioned in most of the general "how to set up linux" documentation, and it would be difficult to set up automatically at install time (although not impossible in theory, if you have a modem). 2: If you check 'man clock' you'll get clock(3), which is not what you want. (try 'man 8 clock'). 3: Most people don't seem to care what time it is anyway. 4: Those few who do care often want to use the xntpd package from louie.udel.edu to sync to an external time source, such as a network time server or radio clock. This mini-HOWTO describes the low-tech approach. If you're at all interested in this sort of thing, I highly recommend that you spend some time at http://www.eecis.udel.edu/~ntp/ which includes all kinds of interesting stuff, including complete info on xntpd and links to NIST and USNO (I have a few more comments on xntpd at the end.) NOTE: If you run more than one OS on your machine, you should only let one of them reset the CMOS clock, so they don't confuse each other. If you regularly run both Linux and Windows on the same machine, you may want to check out some of the shareware clock programs that are available for Windows instead (follow the links from the URL above). Using the 'clock' program: =========================== Everything you need to know is in the clock(8) man page, but this mini-HOWTO will walk you through the process. NOTE: You must be root to run 'clock', or any other program that affects either the system time or the CMOS clock. Checking your installation: =========================== Check your system startup files for a command like 'clock -a' or 'clock -ua'. Depending on which distribution you're using, it might be in /etc/rc.local, or /etc/rc.d/rc.sysinit, or some similar place. If it says 'clock -s' or 'clock -us', change the 's' to an 'a', and then check to see if you have the file /etc/adjtime, which contains a single line that looks something like this: 0.000000 842214901 0.000000 These numbers are the correction factor (in seconds per day), the time the clock was last corrected (in seconds since Jan 1, 1970), and the partial second that was rounded off last time. If you don't have this file, login as root and create it, with a single line that looks like this (all zeros): 0.0 0 0.0 Then run 'clock -a' or 'clock -ua' manually from the shell to update the 2nd number (use the 'u' if your clock is set to Universal instead of local time). Measuring your clock's drift rate: ================================== First, you need to know what time it is :-). Your local time of day number may or may not be accurate. My favorite method is to call WWV's voice announcment at (303)499-7111 (this is a toll call). If you have access to a network time server, you can use the 'ntpdate' program from the xntpd package (use the -b flag to keep the kernel from messing with the CMOS clock). Otherwise use 'date -s hh:mm:ss' to set the kernel time by hand, and then 'clock -w' to set the CMOS clock from the kernel clock. You'll need to remember when you last set the clock, so write down the date someplace where you won't lose it. If you used ntpdate, do 'date +%s' and write down the number of seconds since Jan 1,1970. Then come back some days or weeks later and see how far the clock has drifted. If you're setting the clock by hand, I'd recommend waiting at least two weeks, and only calculate the drift rate to the nearest .1 sec/day. After several months you could get to the nearest .01 sec/day (some people claim more accuracy than that, but I'm being conservative here). If you use 'ntpdate' you shouldn't have to wait that long, but in any case you can always fine-tune it later. You can have cron run 'clock -a' at regular intervals to keep the system time in line with the (corrected) CMOS time. This command will also be run from your startup file every time you boot the system, so if you do that often (as some of us do), that may be enough for your purposes. Note that certain programs may complain if the system time jumps by more than one second at a time, or if it jumps backwards. If you have this problem, you can use xntpd or ntpdate to correct the time more gradually. Example: ======== To Set time: ------------ Login as root. Dial (303)499-7111 (voice), listen for time announcement. Then type: date -s hh:mm:ss but don't press enter until you hear the beep. [You could use 'ntpdate' here instead of 'date', and skip the phone call] This sets the "kernel time". Then type: clock -w This sets the CMOS time to match the kernel time. Then type: date +%j [or 'date +%s' if you used 'ntpdate' instead of 'date' above] and write down the number it gives you for next time. To Reset Time and Check Drift Rate: ---------------------------------- Find the date you wrote down last time. Login as root Then type: clock -a This sets the kernel time to match the current CMOS time. Dial (303)499-7111 (voice), listen for announcement. Then type: date and press enter when you hear the beep, but while you're waiting, write down the time they announce, and don't hang up yet. This tells you what time your machine thought it was, when it should have been exactly on the minute. Now type in date hh:mm:00 using the minute *after* the one that was just announced, and press enter when you hear the beep again (now you can hang up). For hh use the local hour. This sets the "kernel time". Then type: clock -w which writes the new (correct) time to the CMOS clock. Now type: date +%j [or 'date +%s' if that's what you used before] You now have three numbers (two dates and a time) that will allow you to calculate the drift rate. Calculating the Correction Factor: ---------------------------------- When you ran 'date' on the minute, was your machine slow or fast? If it was fast, you'll have to subtract some number of seconds, so write it down as a negative number. If it was slow, you have to add some seconds, so write it down as positive. Now subtract the two dates. If you used 'date +%j', the numbers represent the day-of-year (1-365, or 1-366 in leap years). If you've passed Jan 1 since you last set the clock you'll have to add 365 (or 366) to the 2nd number. If you used 'date +%s' then your number is in seconds, and you'll have to divide it by 86400 to get days. If you already had a correction factor in /etc/adjtime, you'll have to account for the number of seconds you've already corrected. If you've overcorrected, this number will have the opposite sign of the one you just measured; if you've undercorrected it will have the same sign. Multiply the old correction factor by the number of days, and then add the new number of seconds (signed addition-- if the two numbers have the same sign, you'll get a larger number, if they have opposite signs you'll get a smaller number). Then divide the total number of seconds by the number of days to get the new correction factor, and put it in /etc/adjtime in place of the old one. Write down the new date (in seconds or days) for next time. Here's what my /etc/adjtime looks like: -9.600000 845082716 -0.250655 [note 9.6 seconds per day is nearly five minutes per month!] A few words about xntpd: ======================== Your system actually has two clocks-- the battery powered "real time clock" that keeps track of time when the system is turned off (also known as the "CMOS clock", "Hardware clock", or "RTC") and the "kernel time" (sometimes called the "software clock" or "system clock") which is based on the timer interrupt and is initialized from the CMOS clock at boot time. The two will drift at different rates, so they will gradually drift apart from each other, and also away from the "real" time. All references to "the clock" in the xntpd documentation refer to the "kernel clock". When you run xntpd or timed (or any other program that uses the 'adjtimex' system call), the linux kernel assumes that the kernel clock is more accurate than the CMOS clock, and resets the CMOS time every 11 minutes from then on (until you reboot the machine). This means that 'clock' no longer knows when the CMOS clock was last reset, so you can't use the correction factor in /etc/adjtime. You can use 'ntpdate' in your startup file to initially set the clock from a timeserver before starting xntpd. If you don't always have access to an accurate time source when you boot the machine, this can be a bit awkward-- xntpd isn't really designed to be used in situations like that. Xntpd includes drivers for many radio clocks, and can also be set to call NIST's dial-up time service at regular intervals (be sure to calculate the effect on your phone bill when setting the interval between calls). It can also apply a correction factor to the kernel clock if it loses contact with its other sources for an extended period of time. Most radio clocks cost $3-4K, but you can get plans for an inexpensive 'gadget box' (actually a 300 baud modem) that sits between your computer and any shortwave radio tuned to Canada's CHU time station (see ftp://ftp.udel.edu/pub/ntp/gadget.tar.Z). The Heathkit WWV receiver (the "Most Accurate Clock") is also still available (although not as a kit), and costs around $4-500. GPS signals also contain time information, and some GPS recievers can connect to a serial port. This may become the low cost solution in the near future. In theory, someone could write a program to use NIST's dial up time service to calculate the drift rate of both the CMOS clock and the kernel clock automatically. I am not aware of any stand-alone program to do this, but most of the code could probably be borrowed from xntpd.