Re: [ntp:questions] Fudge time offset on client/peer?
Rich Wales wrote: you might experiment with the new interleaved mode to see how it handles the asymmetry. You could peer your home and work refclock ntpds . . . . Actually, I tried that some time back, but it didn't help a bit. I eventually came to understand that xleave mode is designed to reduce the effect of processing delays inside a server's network stack. Sadly, it doesn't do a thing to alleviate asymmetries. Indeed, my current impression (someone please correct me if I'm wrong!) is that NTP is *inherently incapable* of measuring (or even detecting) network latency asymmetries -- the NTP protocol assumes that traffic comes and goes with equal speed, and it has no way to detect situations in which this is not true. That is correct, unfortunately. Noone has been able to come up with a scheme that will allow NTP, or any other program for that matter, to be able to measure any asymmetry. All it can tell is when it sent out a packet and when it got it back. There is nothing that either end can do to give it any indication of the nature of the asymmetry. Neither photons nor electrons can tell you how fast they were moving even in a single segment never mind over multiple segments. I have no idea why anyone would think the xleave would make any difference since it cannot tell you anything about asymmetry. The fudge command is really meant for refclocks and not servers or peers. I don't think that it can be used to express asymmetry for a server or peer. Danny -- This message has been scanned for viruses and dangerous content by MailScanner, and is believed to be clean. ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
Dave, Can we not just introduce an option on the server line for this? That would effectively give you the association. The caviat here is how do you know what to put in the argument? Danny David Mills wrote: Rich, I have the same situation you have, but a dedicated ISDN line and routers that have a mostly symmetric delays. A per-association fudge is not possilbe unless the peer mobilization code is overhauled. There is in fact a calibration mechanism designed to compensate for small inconsistencies using the PPS signal as the ultimate reference. That is controlled by the enable/disable calibrate command and applies to all associations. A command might be introduced that could affect a specified association, but it would have to be given via ntpq after mobilization. Dave Rich Wales wrote: I suggest you don't want that. What you need is a fudge on the interface, not the association. In this situation, I think I really do want a fudge on the association. Consider the issue from the POV of my work desktop. My work desktop has a single network interface, connected to a conventional 100BASE-TX ethernet network. It has fast (and, for my purposes, sufficiently symmetric) connectivity over my school's campus network to stratum-1 and stratum-2 servers run by the campus IT services. In order for my work desktop to see the stratum-1 server I'm running at my home, however, it has to go over the campus network to the cable-modem network servicing the townhouse complex where I live. As I previously mentioned, this cable modem network appears to have an asymmetry, which I would like to fudge away for the benefit of my work desktop (but *not* for my home LAN). If I were to fudge the network interface of my work desktop, this would presumably affect not only its view of my home stratum-1 server, but also my work desktop's view of the campus tickers. What I think I want/need is a way to fudge my work desktop's view of one peer/server, but not another peer/server. That's why I wanted to be able to fudge an association. Rich Wales / ri...@richw.org ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions -- This message has been scanned for viruses and dangerous content by MailScanner, and is believed to be clean. ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
Danny, Sure you can in a couple of lines of code. However, where are you going to put the result? The proto_config() call has a fixed number of fields all tied up with data structures used for name resolution and for remote configuration. The original plan, now only part way completed, was to save everything in the parse ttree while the resolver is out to lunch, then pick up where it left off when lunch is over. That would completely resolve (no pun) the issue and allow an indefinate expansion in configuration options. Dave Danny Mayer wrote: Dave, Can we not just introduce an option on the server line for this? That would effectively give you the association. The caviat here is how do you know what to put in the argument? Danny David Mills wrote: Rich, I have the same situation you have, but a dedicated ISDN line and routers that have a mostly symmetric delays. A per-association fudge is not possilbe unless the peer mobilization code is overhauled. There is in fact a calibration mechanism designed to compensate for small inconsistencies using the PPS signal as the ultimate reference. That is controlled by the enable/disable calibrate command and applies to all associations. A command might be introduced that could affect a specified association, but it would have to be given via ntpq after mobilization. Dave Rich Wales wrote: I suggest you don't want that. What you need is a fudge on the interface, not the association. In this situation, I think I really do want a fudge on the association. Consider the issue from the POV of my work desktop. My work desktop has a single network interface, connected to a conventional 100BASE-TX ethernet network. It has fast (and, for my purposes, sufficiently symmetric) connectivity over my school's campus network to stratum-1 and stratum-2 servers run by the campus IT services. In order for my work desktop to see the stratum-1 server I'm running at my home, however, it has to go over the campus network to the cable-modem network servicing the townhouse complex where I live. As I previously mentioned, this cable modem network appears to have an asymmetry, which I would like to fudge away for the benefit of my work desktop (but *not* for my home LAN). If I were to fudge the network interface of my work desktop, this would presumably affect not only its view of my home stratum-1 server, but also my work desktop's view of the campus tickers. What I think I want/need is a way to fudge my work desktop's view of one peer/server, but not another peer/server. That's why I wanted to be able to fudge an association. Rich Wales / ri...@richw.org ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
Rich, I suggest you don't want that. What you need is a fudge on the interface, not the association. A fudge on the association is hard due to the limitations imposed by the configuration process. For instance, some Sun machines show asymmetric delays on the Ether interfaces of over two milliseconds. It gets even harder with multiple interfaces of different types on the same machine. Even harder for gigabit Etherfaces with interrupt coalescing. Dave Rich Wales wrote: My home network is connected to the outside world via a cable modem. I'm running a refclock at home (on a FreeBSD server running ntpd 4.2.5p215), which I want to refer to both at home and elsewhere. The cable modem system appears to be introducing an offset of about 2 msec (compared to a refclock on my work network, which I have reason to believe is trustworthy). I would like to be able to tell ntpd (4.2.5p181) on my work machine to fudge the offset it gets from my home clock by about 2 msec, in order to counter the cable modem assymetry. However, I do *not* want to fudge the offset on the refclock itself, because I want the hosts on my home network to see the server as is. I'm not sure what to do, because as best I can tell, I can specify a fudge time1 only in connection with a refclock. Is there any way to fudge the offset on a peer or (non-refclock) server configuration command? If not, can this be considered as an enhancement? Or am I missing some very good reason why this is a Really Bad Idea that should (and will) never be implemented? Rich Wales / ri...@richw.org ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
Dave, Interleaved mode doesn't help with asymmetric delays; however, you may notice an ntpq billboard variable xleave, in any mode. The choice of name is probably misleading and might be changed. What this measures is the time from completing all packet fields other than the MAC to the time of the output device interrupt, properly called the transmit drivestamp. In interleaved symmetric and broadcast modes, the transmit drivestamp this is used as the transmit timestamp sent in the following packet so the delays due to output queuing, buffering and crypto operations can be compensated. For noninterleaved modes the variable could be used to back time the transmit timestamp in the next packet. This could produce more jitter, but in the average better time. Dave Dave Hart wrote: Hi Rich, As you've noticed, fudging the offset it's available except for refclocks. However, you might experiment with the new interleaved mode to see how it handles the asymmetry. You could peer your home and work refclock ntpds using something like: peer other maxpoll 8 xleave The big downside to xleave is it is only available for symmetric (peer) associations which are configured explicitly on both sides. Cheers, Dave Hart ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
I suggest you don't want that. What you need is a fudge on the interface, not the association. In this situation, I think I really do want a fudge on the association. Consider the issue from the POV of my work desktop. My work desktop has a single network interface, connected to a conventional 100BASE-TX ethernet network. It has fast (and, for my purposes, sufficiently symmetric) connectivity over my school's campus network to stratum-1 and stratum-2 servers run by the campus IT services. In order for my work desktop to see the stratum-1 server I'm running at my home, however, it has to go over the campus network to the cable-modem network servicing the townhouse complex where I live. As I previously mentioned, this cable modem network appears to have an asymmetry, which I would like to fudge away for the benefit of my work desktop (but *not* for my home LAN). If I were to fudge the network interface of my work desktop, this would presumably affect not only its view of my home stratum-1 server, but also my work desktop's view of the campus tickers. What I think I want/need is a way to fudge my work desktop's view of one peer/server, but not another peer/server. That's why I wanted to be able to fudge an association. Rich Wales / ri...@richw.org ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
Rich, I have the same situation you have, but a dedicated ISDN line and routers that have a mostly symmetric delays. A per-association fudge is not possilbe unless the peer mobilization code is overhauled. There is in fact a calibration mechanism designed to compensate for small inconsistencies using the PPS signal as the ultimate reference. That is controlled by the enable/disable calibrate command and applies to all associations. A command might be introduced that could affect a specified association, but it would have to be given via ntpq after mobilization. Dave Rich Wales wrote: I suggest you don't want that. What you need is a fudge on the interface, not the association. In this situation, I think I really do want a fudge on the association. Consider the issue from the POV of my work desktop. My work desktop has a single network interface, connected to a conventional 100BASE-TX ethernet network. It has fast (and, for my purposes, sufficiently symmetric) connectivity over my school's campus network to stratum-1 and stratum-2 servers run by the campus IT services. In order for my work desktop to see the stratum-1 server I'm running at my home, however, it has to go over the campus network to the cable-modem network servicing the townhouse complex where I live. As I previously mentioned, this cable modem network appears to have an asymmetry, which I would like to fudge away for the benefit of my work desktop (but *not* for my home LAN). If I were to fudge the network interface of my work desktop, this would presumably affect not only its view of my home stratum-1 server, but also my work desktop's view of the campus tickers. What I think I want/need is a way to fudge my work desktop's view of one peer/server, but not another peer/server. That's why I wanted to be able to fudge an association. Rich Wales / ri...@richw.org ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
A per-association fudge is not possible unless the peer mobilization code is overhauled. . . . A command might be introduced that could affect a specified association, but it would have to be given via ntpq after mobilization. Suppose I were to move the refclock to my (dual-homed) firewall -- where (in theory) a per-interface fudge could be specified for the interface to my cable modem (but *not* for the interface to my LAN). Would that make it more feasible to do what I need? Or, alternatively, could some sort of filter application be devised that would sit on my firewall (currently a Linux box) and fudge the NTP packets going between my stratum-1 server and the outside? (Yes, I agree this could easily end up being a messy kludge, but . . . .) Rich Wales / ri...@richw.org ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
Hi Rich, As you've noticed, fudging the offset it's available except for refclocks. However, you might experiment with the new interleaved mode to see how it handles the asymmetry. You could peer your home and work refclock ntpds using something like: peer other maxpoll 8 xleave The big downside to xleave is it is only available for symmetric (peer) associations which are configured explicitly on both sides. Cheers, Dave Hart ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
Re: [ntp:questions] Fudge time offset on client/peer?
you might experiment with the new interleaved mode to see how it handles the asymmetry. You could peer your home and work refclock ntpds . . . . Actually, I tried that some time back, but it didn't help a bit. I eventually came to understand that xleave mode is designed to reduce the effect of processing delays inside a server's network stack. Sadly, it doesn't do a thing to alleviate asymmetries. Indeed, my current impression (someone please correct me if I'm wrong!) is that NTP is *inherently incapable* of measuring (or even detecting) network latency asymmetries -- the NTP protocol assumes that traffic comes and goes with equal speed, and it has no way to detect situations in which this is not true. -- Rich Wales / ri...@richw.org / ri...@stanford.edu Wikipedia: http://en.wikipedia.org/wiki/User:Richwales Facebook: http://www.facebook.com/richwales ___ questions mailing list questions@lists.ntp.org https://lists.ntp.org/mailman/listinfo/questions
