Re: [ntp:questions] Which version of Linux works best?
unruh wrote: On 2010-03-11, Hal Murray hal-use...@ip-64-139-1-69.sjc.megapath.net wrote: Modern Linux kernels don't support PPS in the sense of RFC-whateveritis. There is support for an ioctl that says wake me up when a modem signal changes. gpsd uses that to provide PPS support. I don't have any data. I believe but am not sure, that that uses an interrupt. I think so. But the point is that with the PPS support, the kernel grabs a timestamp in the interrupt routine. The ioctl So? The interrupt still takes the same time to be activated. On a GHZ system, there is enough time in 1usec to run 1000 commands, and it is hard to imagine that many being used to return the ioctl. I have worried That's 1000 machine cycles, not 1000 instructions. On modern systems, I'm not sure that 1000 cycles isn't a typical time for a system call on modern, high level language progammed, bloatware. (I seem to remember hand coding an ISR in assembler to a budget of 100 instructions (for 68000) and it not being that easy.) ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David Woolley wrote: So? The interrupt still takes the same time to be activated. On a GHZ system, there is enough time in 1usec to run 1000 commands, and it is hard to imagine that many being used to return the ioctl. I have worried That's 1000 machine cycles, not 1000 instructions. On modern systems, I'm not sure that 1000 cycles isn't a typical time for a system call on modern, high level language progammed, bloatware. (I seem to remember hand coding an ISR in assembler to a budget of 100 instructions (for 68000) and it not being that easy.) Interrupt times are quite often in the multi-K cycle count for modern operating systems. :-( For ntp it is the variability that really counts, a constant response time will just lead to a small but fixed offset from the true time. If you really care you can fudge away that offset. :-) Terje -- - Terje.Mathisen at tmsw.no almost all programming can be viewed as an exercise in caching ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David Woolley wrote: That's 1000 machine cycles, not 1000 instructions. On modern systems, I'm not sure that 1000 cycles isn't a typical time for a system call on modern, high level language progammed, bloatware. (I seem to remember hand coding an ISR in assembler to a budget of 100 instructions (for 68000) and it not being that easy.) 100 68k instructions will take from 1000 to 5000 clk cycles on the basic 68k. cache, pipelines and superscalar execution started to have increasing impact with 68030 and 68040 ( in exchange killing deterministic reaction times for good ) uwe ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-12, David Woolley da...@ex.djwhome.demon.invalid wrote: unruh wrote: On 2010-03-11, Hal Murray hal-use...@ip-64-139-1-69.sjc.megapath.net wrote: Modern Linux kernels don't support PPS in the sense of RFC-whateveritis. There is support for an ioctl that says wake me up when a modem signal changes. gpsd uses that to provide PPS support. I don't have any data. I believe but am not sure, that that uses an interrupt. I think so. But the point is that with the PPS support, the kernel grabs a timestamp in the interrupt routine. The ioctl So? The interrupt still takes the same time to be activated. On a GHZ system, there is enough time in 1usec to run 1000 commands, and it is hard to imagine that many being used to return the ioctl. I have worried That's 1000 machine cycles, not 1000 instructions. On modern systems, And since most modern processors are pipelined and parallelized it may mean more than 1000 instructions. I'm not sure that 1000 cycles isn't a typical time for a system call on modern, high level language progammed, bloatware. (I seem to remember hand coding an ISR in assembler to a budget of 100 instructions (for 68000) and it not being that easy.) No idea, which is why I would love to see tests to see how long it takes the serial port to respond. I know the parallel port takes something like 1 -2 usec between Timestamp raise parallel port out line get and process interrupt and deliver to kernel interrupt processing module Timestamp (The out line is connected to the parallel port interrupt control line) ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
unruh wrote: On 2010-03-12, David Woolley da...@ex.djwhome.demon.invalid wrote: unruh wrote: On 2010-03-11, Hal Murray hal-use...@ip-64-139-1-69.sjc.megapath.net wrote: Modern Linux kernels don't support PPS in the sense of RFC-whateveritis. There is support for an ioctl that says wake me up when a modem signal changes. gpsd uses that to provide PPS support. I don't have any data. I believe but am not sure, that that uses an interrupt. I think so. But the point is that with the PPS support, the kernel grabs a timestamp in the interrupt routine. The ioctl So? The interrupt still takes the same time to be activated. On a GHZ system, there is enough time in 1usec to run 1000 commands, and it is hard to imagine that many being used to return the ioctl. I have worried That's 1000 machine cycles, not 1000 instructions. On modern systems, And since most modern processors are pipelined and parallelized it may mean more than 1000 instructions. I'm not sure that 1000 cycles isn't a typical time for a system call on modern, high level language progammed, bloatware. (I seem to remember hand coding an ISR in assembler to a budget of 100 instructions (for 68000) and it not being that easy.) No idea, which is why I would love to see tests to see how long it takes the serial port to respond. I know the parallel port takes something like 1 -2 usec between Timestamp raise parallel port out line get and process interrupt and deliver to kernel interrupt processing module Timestamp (The out line is connected to the parallel port interrupt control line) Afair there used to be response data ( delay, jitter) available for RT-Linux but I can't find it at the moment. ( and it is different for an ISA connected Interface and a PCI connected one.) you have the aliasing jitter between incoming serial signal and the sampling clock to reckon with. ( bittime divided by (over)sampling rate 4,8,32,64? in the receiver, only have the data for the Z8530 / Z8035 type parts. ) uwe ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
unruh wrote: On 2010-03-12, Terje Mathisenterje.mathisen at tmsw.no wrote: OTOH, I have personally never seen this on any of my S1 servers which all use the serial port. Not sure how you would see that. If the interrupt were delayed by one ms ntp would not know. It would see something only if that delay were variable. A gated interrupt would show up as high jitter, since the delays would follow a sawtooth curve, depending upon when in the cycle the external interrupt occured. Terje -- - Terje.Mathisen at tmsw.no almost all programming can be viewed as an exercise in caching ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-12, Uwe Klein uwe_klein_habertw...@t-online.de wrote: unruh wrote: On 2010-03-12, David Woolley da...@ex.djwhome.demon.invalid wrote: unruh wrote: On 2010-03-11, Hal Murray hal-use...@ip-64-139-1-69.sjc.megapath.net wrote: Modern Linux kernels don't support PPS in the sense of RFC-whateveritis. There is support for an ioctl that says wake me up when a modem signal changes. gpsd uses that to provide PPS support. I don't have any data. I believe but am not sure, that that uses an interrupt. I think so. But the point is that with the PPS support, the kernel grabs a timestamp in the interrupt routine. The ioctl So? The interrupt still takes the same time to be activated. On a GHZ system, there is enough time in 1usec to run 1000 commands, and it is hard to imagine that many being used to return the ioctl. I have worried That's 1000 machine cycles, not 1000 instructions. On modern systems, And since most modern processors are pipelined and parallelized it may mean more than 1000 instructions. I'm not sure that 1000 cycles isn't a typical time for a system call on modern, high level language progammed, bloatware. (I seem to remember hand coding an ISR in assembler to a budget of 100 instructions (for 68000) and it not being that easy.) No idea, which is why I would love to see tests to see how long it takes the serial port to respond. I know the parallel port takes something like 1 -2 usec between Timestamp raise parallel port out line get and process interrupt and deliver to kernel interrupt processing module Timestamp (The out line is connected to the parallel port interrupt control line) Afair there used to be response data ( delay, jitter) available for RT-Linux but I can't find it at the moment. ( and it is different for an ISA connected Interface and a PCI connected one.) you have the aliasing jitter between incoming serial signal and the sampling clock to reckon with. ( bittime divided by (over)sampling rate 4,8,32,64? in the receiver, only have the data for the Z8530 / Z8035 type parts. ) I cerainly would not rely on the data in/out for the interrupt as it might well have clock aliasing. But is there not a specific pin on the serial port which is an immediate interrupt pin like the interrupt pin on the parallel port? uwe ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
unruh wrote: I cerainly would not rely on the data in/out for the interrupt as it might well have clock aliasing. But is there not a specific pin on the serial port which is an immediate interrupt pin like the interrupt pin on the parallel port? The hardware supports interrupt on DCD, DTR, CTS, ( all status bits on change ) http://www.national.com/ds/PC/PC16550D.pdf ( page 18 ) look out on masking and priority. I've only done the bitbanging stuff on the Z8530 family in an embedded app that used HDLC transmission and had to control/get status from a wireless transceiver. The Z8530 can do nearly everything except begging for food. But the datasheets and appnotes are certainly cryptic. uwe uwe ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
John Hasler wrote: David writes: My report to chrony-dev list, along with links to mrtg graphs, never made it to the list, and although I can send that again... Please send it directly to me. OK it was a few months back, I'll try to find it and make sure links to the graphs still work. Also I've been receiving ok from mailing list again from early Feb when you sorted it out - thanks. cheers David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Richard B. Gilbert wrote: Chuck Swiger wrote: On Mar 10, 2010, at 1:05 PM, John Hasler wrote: I gather that crony is intended for machines with infrequent network connections. That was one of the goals when it was first developed ten years ago. It has gone far beyond that now. OK. I can't imagine trying to run it for a permanently networked stratum-1 timesource. Why? I've seen monitoring data from the NTP pool project for people using other NTP implementations, and they don't seem to be nearly as reliable timesources as the original ntpd implementation. It's not just my opinion: http://www.pool.ntp.org/en/join/configuration.html Use the standard ntpd We are all for software diversity, but a significant percentage of the it's not working questions that come in are for software other than ntpd. You can use the pool with any program speaking NTP, but if you are going to join the pool we recommend you use ntpd. Can you give me a pointer to some IPs in the NTP pool using crony, so we can check their scores at http://www.pool.ntp.org/scores/IP ...? Regards, I believe that the servers in the NTP pool are ALL using NTPD. Chrony is an entirely separate product unrelated to NTPD except for the fact that it does something vaguely similar. In 1997, before I'd used ntpd I had chrony on a pair of systems used for dialup connections peered together. Later, ntpd on my servers worked without any problem using the two chrony sources. I've not been able to successfully peer between chrony and ntpd though, but I don't really have spare hardware at moment to really test this out. David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On Wed, Mar 10, 2010 at 03:25:45PM -, David J Taylor wrote: - and one for Bill, how much better might chrony be than official NTP? In my experience, chrony is about 3-20 times better than NTP when using the same poll interval. The more stable is temperature and CPU load the smaller is the difference. Does it have a preferred Linux, or even freeBSD? The BSD drivers in chrony are using only adjtime() call for clock corrections (similar to the NTP daemon mode), so for optimal results running chrony on Linux might be necessary. There were some issues related to the kernel tickless feature which affected chrony's initial synchronization. They were fixed only very recently, so if you need to get below 1 us in less than one minute, the latest kernel (2.6.33) or kernel compiled without CONFIG_NO_HZ is recommended. As for PPS source, LinuxPPS patch can be applied to kernel (hopefully it will be merged into mainline soon), or PPS samples from gpsd can be used instead, versions 2.90 and later works best. In my tests the gpsd source has about two times worse dispersion and there is a small shift when compared to the kernel source. There is also a difference in resolution (nanoseconds vs microseconds). The system would /not/ be in a temperature controlled environment. I did a NTP vs chrony comparison last June with GPS 18x LVC in an office environment, clock drift was moving in about 0.8ppm range. Here are distributions of PPS samples received from gpsd: http://fedorapeople.org/~mlichvar/chrony/chrony_vs_ntp.png With recent chrony, NTP and kernel versions the results might be different though. -- Miroslav Lichvar ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On Thu, Mar 11, 2010 at 11:53:28AM +, David Lord wrote: In 1997, before I'd used ntpd I had chrony on a pair of systems used for dialup connections peered together. Later, ntpd on my servers worked without any problem using the two chrony sources. I've not been able to successfully peer between chrony and ntpd though, but I don't really have spare hardware at moment to really test this out. As chrony supports only version 3 of the NTP protocol, you might need to add version 3 to the chrony peer specification in ntp.conf. -- Miroslav Lichvar ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Miroslav Lichvar mlich...@redhat.com wrote in message news:20100311124036.ga22...@localhost... [] I did a NTP vs chrony comparison last June with GPS 18x LVC in an office environment, clock drift was moving in about 0.8ppm range. Here are distributions of PPS samples received from gpsd: http://fedorapeople.org/~mlichvar/chrony/chrony_vs_ntp.png With recent chrony, NTP and kernel versions the results might be different though. -- Miroslav Lichvar Miroslav, Thanks for that, a most interesting comparison, and thanks to everyone for their input. I've recently switched the old (2005) FreeBSD system back on, to see how well in performs in my own environment. From what's been said, I rather suspect that were I to go for a more modern, faster, Intel Atom system, any improvement in accuracy I might get could be swamped by the temperature changes in the room. There's also, I will admit, a slight doubt about the effort involved for the benefit to be gained. With Windows, I am quite happy, and configuring, using or testing NTP is no problem. With FreeBSD is seems that the old PPS atom driver has gone, and I may need to configure yet another driver - gpsd. The number of variants of Linux doesn't help - I only need a command-line or Telnet interface. And remembering how long it took to recompile the kernel last time, and the amount of help I needed to know how to do that, also fills me with doubt. So I suspect that the performance I'm now seeing from Windows (well within 100us) may well be good enough for me. Perhaps if I get more free time, and a little more income this year, I may get a paperback-sized Intel Atom box and see how it does. At least some do have serial ports! And I would be most interested to hear of anyone who does configure such a device. Cheers, David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Miroslav Lichvar wrote: On Thu, Mar 11, 2010 at 11:53:28AM +, David Lord wrote: In 1997, before I'd used ntpd I had chrony on a pair of systems used for dialup connections peered together. Later, ntpd on my servers worked without any problem using the two chrony sources. I've not been able to successfully peer between chrony and ntpd though, but I don't really have spare hardware at moment to really test this out. As chrony supports only version 3 of the NTP protocol, you might need to add version 3 to the chrony peer specification in ntp.conf. Thanks for that as it would explain why I believed it used to work ok. I'll add a note to ntp.conf of servers for when I get round to trying again. cheers David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David J Taylor wrote: Miroslav Lichvar mlich...@redhat.com wrote in message news:20100311124036.ga22...@localhost... [] I did a NTP vs chrony comparison last June with GPS 18x LVC in an office environment, clock drift was moving in about 0.8ppm range. Here are distributions of PPS samples received from gpsd: http://fedorapeople.org/~mlichvar/chrony/chrony_vs_ntp.png With recent chrony, NTP and kernel versions the results might be different though. -- Miroslav Lichvar Miroslav, Thanks for that, a most interesting comparison, and thanks to everyone for their input. I've recently switched the old (2005) FreeBSD system back on, to see how well in performs in my own environment. From what's been said, I rather suspect that were I to go for a more modern, faster, Intel Atom system, any improvement in accuracy I might get could be swamped by the temperature changes in the room. There's also, I will admit, a slight doubt about the effort involved for the benefit to be gained. With Windows, I am quite happy, and configuring, using or testing NTP is no problem. With FreeBSD is seems that the old PPS atom driver has gone, and I may need to configure yet another driver - gpsd. The number of variants of Linux doesn't help - I only need a command-line or Telnet interface. And remembering how long it took to recompile the kernel last time, and the amount of help I needed to know how to do that, also fills me with doubt. So I suspect that the performance I'm now seeing from Windows (well within 100us) may well be good enough for me. Perhaps if I get more free time, and a little more income this year, I may get a paperback-sized Intel Atom box and see how it does. At least some do have serial ports! And I would be most interested to hear of anyone who does configure such a device. I'd have no hesitation trying any of Linux, Free or NetBSD on an atom board. Options in bios and kernel regarding power saving etc are main concerns and would need to be disabled. For gps refclock via serial port I'd be happy to risk NetBSD although only atom system I have is eeepc without serial port. With NetBSD and via c3 with gps via serial port I really need to be using nanosecond rather than microsecond scales on mrtg graphs. Lots of Linux distributions I've tried come complete with all bells, whistles and kitchen sink, but for using other than packaged system are to me more difficult to maintain than very basic Slackware or similar. From what's been said in thread so far, there are advantages to using Linux if you want to experiment. At least for me there is very little difference between install of Slackware, FreeBSD or NetBSD and then manual config of ntp.conf (major decision is what editor to use, joe, nano or vi). cheers David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Once upon a time, Miroslav Lichvar mlich...@redhat.com said: As for PPS source, LinuxPPS patch can be applied to kernel (hopefully it will be merged into mainline soon), or PPS samples from gpsd can be used instead, versions 2.90 and later works best. There's also the user-space shmpps (which is packaged in Fedora, so yum install shmpps will get it; edit /etc/sysconfig/shmpps to configure). -- Chris Adams cmad...@hiwaay.net Systems and Network Administrator - HiWAAY Internet Services I don't speak for anybody but myself - that's enough trouble. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David J Taylor wrote: Miroslav Lichvar mlich...@redhat.com wrote in message news:20100311124036.ga22...@localhost... [] I did a NTP vs chrony comparison last June with GPS 18x LVC in an office environment, clock drift was moving in about 0.8ppm range. Here are distributions of PPS samples received from gpsd: http://fedorapeople.org/~mlichvar/chrony/chrony_vs_ntp.png With recent chrony, NTP and kernel versions the results might be different though. -- Miroslav Lichvar Miroslav, Thanks for that, a most interesting comparison, and thanks to everyone for their input. I've recently switched the old (2005) FreeBSD system back on, to see how well in performs in my own environment. From what's been said, I rather suspect that were I to go for a more modern, faster, Intel Atom system, any improvement in accuracy I might get could be swamped by the temperature changes in the room. There's also, I will admit, a slight doubt about the effort involved for the benefit to be gained. With Windows, I am quite happy, and configuring, using or testing NTP is no problem. With FreeBSD is seems that the old PPS atom driver has gone, and I may need to configure yet another driver - gpsd. The number of variants of Linux doesn't help - I only need a command-line or Telnet interface. And remembering how long it took to recompile the kernel last time, and the amount of help I needed to know how to do that, also fills me with doubt. So I suspect that the performance I'm now seeing from Windows (well within 100us) may well be good enough for me. Perhaps if I get more free time, and a little more income this year, I may get a paperback-sized Intel Atom box and see how it does. At least some do have serial ports! And I would be most interested to hear of anyone who does configure such a device. Cheers, David 100 microseconds is pretty good. Getting the time *into* a computer takes time and the time taken is not easy to measure. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? The system would /not/ be in a temperature controlled environment. Thanks, David And now for something completely different: Build your own ( possibly cross platform ) mini/embedded/single purpose linux distribution: http://en.opensuse.org/Build_Service uwe ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Uwe Klein wrote: And now for something completely different: Build your own ( possibly cross platform ) mini/embedded/single purpose linux distribution: http://en.opensuse.org/Build_Service uwe forgot this: http://susestudio.com/ ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-11, Richard B. Gilbert rgilber...@comcast.net wrote: David J Taylor wrote: Miroslav Lichvar mlich...@redhat.com wrote in message news:20100311124036.ga22...@localhost... [] I did a NTP vs chrony comparison last June with GPS 18x LVC in an office environment, clock drift was moving in about 0.8ppm range. Here are distributions of PPS samples received from gpsd: http://fedorapeople.org/~mlichvar/chrony/chrony_vs_ntp.png With recent chrony, NTP and kernel versions the results might be different though. -- Miroslav Lichvar Miroslav, Thanks for that, a most interesting comparison, and thanks to everyone for their input. I've recently switched the old (2005) FreeBSD system back on, to see how well in performs in my own environment. From what's been said, I rather suspect that were I to go for a more modern, faster, Intel Atom system, any improvement in accuracy I might get could be swamped by the temperature changes in the room. There's also, I will admit, a slight doubt about the effort involved for the benefit to be gained. With Windows, I am quite happy, and configuring, using or testing NTP is no problem. With FreeBSD is seems that the old PPS atom driver has gone, and I may need to configure yet another driver - gpsd. The number of variants of Linux doesn't help - I only need a command-line or Telnet interface. And remembering how long it took to recompile the kernel last time, and the amount of help I needed to know how to do that, also fills me with doubt. So I suspect that the performance I'm now seeing from Windows (well within 100us) may well be good enough for me. Perhaps if I get more free time, and a little more income this year, I may get a paperback-sized Intel Atom box and see how it does. At least some do have serial ports! And I would be most interested to hear of anyone who does configure such a device. Cheers, David 100 microseconds is pretty good. Getting the time *into* a computer takes time and the time taken is not easy to measure. Considering that chrony can give sub microsecond resolution from say a GPS source, ( and ntpd 2usec) 100usec is good only only in a certain defintion of good. Getting the time into the computer from a refclock is on the 1usec level ( measured), certainly not 100usec. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Modern Linux kernels don't support PPS in the sense of RFC-whateveritis. There is support for an ioctl that says wake me up when a modem signal changes. gpsd uses that to provide PPS support. I don't have any data. I believe but am not sure, that that uses an interrupt. I think so. But the point is that with the PPS support, the kernel grabs a timestamp in the interrupt routine. The ioctl stuff just wakes up the user program so it can grab the timestamp. On a lightly loaded system, that will probably work OK. But if the system gets busy, there will be more noise in the data. -- These are my opinions, not necessarily my employer's. I hate spam. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-11, Hal Murray hal-use...@ip-64-139-1-69.sjc.megapath.net wrote: Modern Linux kernels don't support PPS in the sense of RFC-whateveritis. There is support for an ioctl that says wake me up when a modem signal changes. gpsd uses that to provide PPS support. I don't have any data. I believe but am not sure, that that uses an interrupt. I think so. But the point is that with the PPS support, the kernel grabs a timestamp in the interrupt routine. The ioctl So? The interrupt still takes the same time to be activated. On a GHZ system, there is enough time in 1usec to run 1000 commands, and it is hard to imagine that many being used to return the ioctl. I have worried about that and it would be nice if someone ran the system such that say one timed when the out pin on the parallel port was activated, and the time that the serial port ioctl returned. I know on my parallel port interrupt, the test I ran showed that the time between activating the pin on the parallel port and the parallel port interrupt service routine timestamping the interrupt was of the order of 1usec. It would be nice to see what it is for the serial port. I doubt it is much more than that. stuff just wakes up the user program so it can grab the timestamp. On a lightly loaded system, that will probably work OK. But if the system gets busy, there will be more noise in the data. If the system gets busy, the interrupts themselves will have delays as well, putting noise in the system. One of the reasons I used the parallel port was because the parallel interrupt has a higher priority than the serial (comes earlier in the interrupt chain). But measurements would be great. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
unruh wrote: [snip] time that the serial port ioctl returned. I know on my parallel port interrupt, the test I ran showed that the time between activating the pin on the parallel port and the parallel port interrupt service routine timestamping the interrupt was of the order of 1usec. It would be nice to see what it is for the serial port. I doubt it is much more than that. You might be wrong, and this is supposedly the reason phk started using the parallel port in the first place: At least some serial ports will gate their interrupt signal to the resolution of their internal clock frequency or hw polling interval, which can result in multiple us's of delay. OTOH, I have personally never seen this on any of my S1 servers which all use the serial port. Terje -- - Terje.Mathisen at tmsw.no almost all programming can be viewed as an exercise in caching ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
[ntp:questions] Which version of Linux works best?
Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? The system would /not/ be in a temperature controlled environment. Thanks, David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Once upon a time, David J Taylor david-tay...@blueyonder.co.uk said: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. Fedora 12 i686 was rebuilt with compilter optimization flags set for the Atom CPU. I don't know how much of a difference that might make with NTP, but I figured I'd toss it out as general info. -- Chris Adams cmad...@hiwaay.net Systems and Network Administrator - HiWAAY Internet Services I don't speak for anybody but myself - that's enough trouble. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? I'd use Ubuntu, simply because those systems are very close to 'maintenance-free', i.e. the OS takes care of pretty much all patching/updates/install etc. Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? Sub-us vs 1-5 us, but only with special hw and a non-standard motherboard crystal freq source. - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? The system would /not/ be in a temperature controlled environment. In which case you should probably be happy with sub-10 us performance anyway. Terje Thanks, David -- - Terje.Mathisen at tmsw.no almost all programming can be viewed as an exercise in caching ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? Linux I think is just the kernel whilst filesystem and packages are very diverse between distributions. If you don't already know Linux well already then I'd suggest FreeBSD as being a more solid base than a Linux distribution. Otherwise I favour Slackware/Centos mainly because I'm more familiar with filesystem layout vs that of other distros such as Ubuntu. I have had chrony running on NetBSD but as with ntpd can't get most recent versions to install/run. It should be ok on Linux and probably on FreeBSD but you'd need confirmation vs refclock drivers on FreeBSD for that. Certainly here on NetBSD chrony looks to give offsets a third or less than ntpd and avoids the square wave offset traces from ntpd that result from inability to correct for temperature changes, however I had problems with incompatibility between different chrony versions so given up to it until I have some spare system and time to work with. David Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? The system would /not/ be in a temperature controlled environment. Thanks, David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? The system would /not/ be in a temperature controlled environment. Thanks, David If you're into using development releases of NTP, it's worth noting that there are Debian packages available, so you don't have to compile your own: http://packages.ntp.org/debian/ (Although I don't use Debian Lenny myself, so I can't confirm they're well-maintained or anything.) -- Matt Nordhoff ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On Mar 10, 2010, at 7:25 AM, David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? The hardware you get matters much more than which OS or flavor of Linux you use; tweaking kernel compiler optimization flags matters even less. It would be good to look into the hardware you get in terms of support for HPET, p-state invariant TSC, or how good the ACPI timers are. Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? There are some nice comparisons between different platforms here: http://www.dragonflybsd.org/presentations/nanosleep/ PHK, who wrote the FreeBSD timer code, is measuring ~150 ns timekeeping precision for a stratum-1 timesource: http://phk.freebsd.dk/soekris/pps/ http://phk.freebsd.dk/pubs/timecounter.pdf You might find some threads here interesting, also: http://unix.derkeiler.com/Mailing-Lists/FreeBSD/current/2005-10/1018.html - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? I gather that crony is intended for machines with infrequent network connections. I can't imagine trying to run it for a permanently networked stratum-1 timesource. The system would /not/ be in a temperature controlled environment. That's unfortunate; this effects your time stability more than any other factor being considered. Well, try to make sure you get the PPS timesource working, as that will help. Regards, -- -Chuck ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-10, Matt Nordhoff mnordh...@mattnordhoff.com wrote: If you're into using development releases of NTP, it's worth noting that there are Debian packages available, so you don't have to compile your own: http://packages.ntp.org/debian/ (Although I don't use Debian Lenny myself, so I can't confirm they're well-maintained or anything.) Our auto-builder generates an i386 binary package after each ntp-dev release. Seven of the machines I maintain use these packages to track the ntp-dev releases. Our auto-builder also generates source packages which may be used to build an ntp-dev deb on non-i386 architectures. I use them on my amd64 (Lenny) box. -- Steve Kostecke koste...@ntp.org NTP Public Services Project - http://support.ntp.org/ ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-10, David Lord sn...@lordynet.org wrote: David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? The key issues are whether or not the kernel is power saving. You want to switch that off-- the kernel slowing down and speeding up as the mood takes it. You want to have it running at the same speed always ( what that is does not matter, just that it is the same so the kernel timekeeping does not get messed up.) Any kernel speed is far far higher than needed for ntp-- and old 80486 would probably be sufficient- but changes are a disaster. Since all distros use the same kernel ( with perhaps some mods) and ntp package, there is really not much difference between distros. Some pile on more junk than others. Use what you are familiar with. Linux I think is just the kernel whilst filesystem and packages are very diverse between distributions. If you don't already know Linux well already then I'd suggest FreeBSD as being a more solid base than a Linux distribution. I do not think there is any difference. Otherwise I favour Slackware/Centos mainly because I'm more familiar with filesystem layout vs that of other distros such as Ubuntu. I have had chrony running on NetBSD but as with ntpd can't get most recent versions to install/run. It should be ok on Linux ah. Could you please let the chrony people know what the problem is. Bugs need to be and can be fixed. Have you tried the latest release which came out about 1 month ago? chrony.tuxfamily.org and probably on FreeBSD but you'd need confirmation vs refclock drivers on FreeBSD for that. Certainly here on NetBSD chrony looks to give offsets a third or less than ntpd and avoids the square wave offset traces from ntpd that result from inability to correct for temperature changes, however I had problems with incompatibility between different chrony versions so given up to it until I have some spare system and time to work with. ?? chronyc must be the same version as chronyd. but what do you mean by incompatibility? David Two secondary questions: - how much better might FreeBSD be than Linux? Any actual measurements? - and one for Bill, how much better might chrony be than official NTP? Does it have a preferred Linux, or even freeBSD? No prefered Linux. It does I believe run on BSD, but have never tried it. It has something like a factor of 3 better perfomance than ntpd, primarily because it remembers the past ( ntpd forgets the past) and can use that to imporve its estimates. It is also far far faster to converge (minutes rather than half a day). The system would /not/ be in a temperature controlled environment. chrony corrects for temp variations far better than does ntpd. There are versions of ntpd which use on onboard thermometer to model and compensate for temp variations, which have a much better performance than straight ntpd does ( probably about the same level as chrony's performance, which essentially uses the clock frequency as a temp measurement). Thanks, David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
From: Chuck Swiger cswi...@mac.com Date: Wed, 10 Mar 2010 09:32:21 -0800 Sender: questions-bounces+oberman=es@lists.ntp.org On Mar 10, 2010, at 7:25 AM, David J Taylor wrote: Yes, I know it's one of those low long is a piece of string questions, but I'm now considering a dual-core Intel Atom system, which is Compatible with Linux according the the very minimal blurb I have right now. If the system is to be used purely for NTP with Linux as a serial-port GPS/PPS stratum-1 server (and, yes, I know dual-core isn't needed for that, but I might want to boot Windows-7 64-bit occasionally), and considering that I know very little about Linux, which version of Linux would the group recommend? Does it make any difference as far as timekeeping is concerned? The hardware you get matters much more than which OS or flavor of Linux you use; tweaking kernel compiler optimization flags matters even less. It would be good to look into the hardware you get in terms of support for HPET, p-state invariant TSC, or how good the ACPI timers are. I'd suggest that you simply run NTP on a uniprocessor. Disable any frequency management in the OS. This eliminates most of the major issues impacting NTP and you really don't need more than one old, slow CPU to do the job. Avoid newer, high performance network cards that do interrupt coalescing or be sure it is disabled. It will shoot the jitter between the server and clients through the roof. FWIW, all of my stratum 1 NTP servers are running FreeBSD 7 on P4 uniprocessors. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Chuck Swiger writes: I gather that crony is intended for machines with infrequent network connections. That was one of the goals when it was first developed ten years ago. It has gone far beyond that now. I can't imagine trying to run it for a permanently networked stratum-1 timesource. Why? -- John Hasler jhas...@newsguy.com Dancing Horse Hill Elmwood, WI USA ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On Mar 10, 2010, at 1:05 PM, John Hasler wrote: I gather that crony is intended for machines with infrequent network connections. That was one of the goals when it was first developed ten years ago. It has gone far beyond that now. OK. I can't imagine trying to run it for a permanently networked stratum-1 timesource. Why? I've seen monitoring data from the NTP pool project for people using other NTP implementations, and they don't seem to be nearly as reliable timesources as the original ntpd implementation. It's not just my opinion: http://www.pool.ntp.org/en/join/configuration.html Use the standard ntpd We are all for software diversity, but a significant percentage of the it's not working questions that come in are for software other than ntpd. You can use the pool with any program speaking NTP, but if you are going to join the pool we recommend you use ntpd. Can you give me a pointer to some IPs in the NTP pool using crony, so we can check their scores at http://www.pool.ntp.org/scores/IP ...? Regards, -- -Chuck ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
unruh wrote: On 2010-03-10, David Lord sn...@lordynet.org wrote: If you don't already know Linux well already then I'd suggest FreeBSD as being a more solid base than a Linux distribution. Only lightweight Linux I've tried recently is Slackware which seemed in many ways similar to BSD but required too much effort to get a working desktop with packages I wanted. Ubuntu has near all packages I need but isn't as easy to customise, eg install chrony and ntpd gets removed (I know there will be a way to have both installed at same time). I do not think there is any difference. I have had chrony running on NetBSD but as with ntpd can't get most recent versions to install/run. It should be ok on Linux ah. Could you please let the chrony people know what the problem is. Bugs need to be and can be fixed. Have you tried the latest release which came out about 1 month ago? chrony.tuxfamily.org There seem to be lots of modifications to syscalls etc required for it to work on NetBSD, if I had such a modified version it might compile and run ok (same goes for recent versions of ntpd). Chrony version I tried was 1.24-pre1. That seemed to compile ok but there were problems with name lookup, logging failure via chronyc from different NetBSD system running v1.23, and reported offsets suddenly jumping from 300us to 100ms for short periods. My report to chrony-dev list, along with links to mrtg graphs, never made it to the list, and although I can send that again, my systems are all now setup back to using ntpd. David ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
Chuck Swiger wrote: On Mar 10, 2010, at 1:05 PM, John Hasler wrote: I gather that crony is intended for machines with infrequent network connections. That was one of the goals when it was first developed ten years ago. It has gone far beyond that now. OK. I can't imagine trying to run it for a permanently networked stratum-1 timesource. Why? I've seen monitoring data from the NTP pool project for people using other NTP implementations, and they don't seem to be nearly as reliable timesources as the original ntpd implementation. It's not just my opinion: http://www.pool.ntp.org/en/join/configuration.html Use the standard ntpd We are all for software diversity, but a significant percentage of the it's not working questions that come in are for software other than ntpd. You can use the pool with any program speaking NTP, but if you are going to join the pool we recommend you use ntpd. Can you give me a pointer to some IPs in the NTP pool using crony, so we can check their scores at http://www.pool.ntp.org/scores/IP ...? Regards, I believe that the servers in the NTP pool are ALL using NTPD. Chrony is an entirely separate product unrelated to NTPD except for the fact that it does something vaguely similar. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-10, Richard B. Gilbert rgilber...@comcast.net wrote: Chuck Swiger wrote: On Mar 10, 2010, at 1:05 PM, John Hasler wrote: I gather that crony is intended for machines with infrequent network connections. That was one of the goals when it was first developed ten years ago. It has gone far beyond that now. OK. I can't imagine trying to run it for a permanently networked stratum-1 timesource. Why? I've seen monitoring data from the NTP pool project for people using other NTP implementations, and they don't seem to be nearly as reliable timesources as the original ntpd implementation. It's not just my opinion: Uh, just because alternative X does not work well, does not mean that alternative Y does not as well. Chrony works very well. I run it stably for years. It does a much better job than does ntpd at disciplining the clocks ( roughly a factor of 2 to 3 smaller offsets), which I suspect is because of its far faster response to frequency changes caused eg by temperature changes. http://www.pool.ntp.org/en/join/configuration.html Use the standard ntpd We are all for software diversity, but a significant percentage of the it's not working questions that come in are for software other than ntpd. You can use the pool with any program speaking NTP, but if you are going to join the pool we recommend you use ntpd. Can you give me a pointer to some IPs in the NTP pool using crony, so we can check their scores at http://www.pool.ntp.org/scores/IP ...? ?? How would we know, especially since David Mills says they can detect implimentations like chrony and get them out of the pool. Regards, I believe that the servers in the NTP pool are ALL using NTPD. Chrony is an entirely separate product unrelated to NTPD except for the fact that it does something vaguely similar. If by vaguely similar you mean it disciplines the local clocks on a computer by exchanging ntp datagrams with ntp servers, and responds to ntp queries with ntp datagrams, then yes, it does something vaguely similar. Most would say that it does the same thing as ntpd does ( but better), but like Humpty Dumpty, you are I guess allowed to define your own terms in whatever way you want. ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On Mar 10, 2010, at 4:59 PM, unruh wrote: I've seen monitoring data from the NTP pool project for people using other NTP implementations, and they don't seem to be nearly as reliable timesources as the original ntpd implementation. It's not just my opinion: Uh, just because alternative X does not work well, does not mean that alternative Y does not as well. Chrony works very well. I run it stably for years. It does a much better job than does ntpd at disciplining the clocks ( roughly a factor of 2 to 3 smaller offsets), which I suspect is because of its far faster response to frequency changes caused eg by temperature changes. Ah, that sounds great. I gather from these comments that you have data which you can make available? Can you give me a pointer to some IPs in the NTP pool using crony, so we can check their scores athttp://www.pool.ntp.org/scores/IP ...? ?? How would we know, especially since David Mills says they can detect implimentations like chrony and get them out of the pool. That seems to be a strange thing for David Mills to say, as Ask Hansen is the pool maintainer. Do you have a reference as to where David said such a thing? Regards, -- -Chuck ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions
Re: [ntp:questions] Which version of Linux works best?
On 2010-03-11, Chuck Swiger cswi...@mac.com wrote: On Mar 10, 2010, at 4:59 PM, unruh wrote: ?? How would we know, especially since David Mills says they can detect implimentations like chrony and get them out of the pool. That seems to be a strange thing for David Mills to say, as Ask Hansen is the pool maintainer. Do you have a reference as to where David said such a thing? Based on my quick review ... The pool code appears to use the Perl Net::NTP module to query the individual pool servers (with the same control messages used by ntpq). An incomplete NTP implementation (i.e. one that does not respond to NTP control messages) will never achieve a high enough quality rating to warrant inclusion in the pool zone. -- Steve Kostecke koste...@ntp.org NTP Public Services Project - http://support.ntp.org/ ___ questions mailing list questions@lists.ntp.org http://lists.ntp.org/listinfo/questions