Accuracy of the Network Time Synchronization
Proposal
We are looking for volunteers who would like to participate in the following measurement.
Goal
The exact measurement of the accuracy of time synchronization in a large network.
Importance of the Problem
Active measurement of QoS parameters requires two workstations with synchronized clocks, the first workstation acting as a traffic generator and the second one as a receiver. Similar situation is in the passive measurement when we need two monitoring sites. It is clear that the precision of QoS measurement (e.g. one-way delay) cannot be better than the accuracy of clock synchronization.
The best method of clock synchronization is to use a precise external time source in each node of measurement; however, this solution has several disadvantages: it is expensive, installation is complex, and it is not scalable. Using some method of clock synchronization via network would be advantageous. The main problem is the accuracy of such a method.
Method
Workstation runs ntpd v.4 which is configured as stratum and disciplined by one primary (stratum) server across the network. GPS receiver supplies PPS (Pulse Per Second) signal, which is received as a control signal of asynchronous port. As local time is synchronized, the sub second part of PPS time stamp (given by the local clock in a workstation) can be interpreted as the offset between local time and GPS time and therefore as the actual accuracy of synchronization. The resolution of a described method is in nanosecond order; however, the absolute accuracy is better than 10 microseconds and standard deviation does not exceed 1 microsecond.
The quality of clock synchronization depends on the local quartz oscillator (temperature influence), on the accuracy of the primary ntp server, on the characteristic of network (delay, delay distribution), and on the loop back control in ntp process. The discussion of all of these factors is beyond the scope of this text; however, assuming good primary ntp server over the WAN and slow changes of the temperature of local clock quartz, the accuracy of synchronization depends mainly on the network.
Duration
In order to obtain representative results the measurement should be done for at least several days. Optimal duration is one week for each site.
Requirements
Site of measurement can be located in a PoP or anywhere in connected network. Both variants should be tested. The site consists of
- PC workstation (at least Pentium 120), Linux with nanokernel patch (it is available for 2.2.18, 2.2.19, 2.4.3 and 2.4.16), and stable version of NTP v.4
- network connection to the network which is being tested
- GPS receiver
- primary ntp server located in our pop
- program for measuring of actual time difference and result storing
- limited number of GPS receivers (including all cables and power supply)
Plan
Participant provides a workstation running Linux with
- a) patched kernel, installed and configured ntp.
or- b) root access so that we can do the rest of the installation.
Once the workstation is ready and its clock is synchronized, GPS receiver can be connected and the measurement started.
Links
- first results of our measurement are here
- program for measurement of clock offset is here
- our preliminary work can be found at http://www.cesnet.cz/doc/techzpravy/2001/18/
- QoS measurement tool is described at http://www.cesnet.cz/doc/techzpravy/2001/07/
- nanokernel patch is available at ftp://ftp.kernel.org/pub/linux/daemons/ntp
- home page of NTP is http://www.ntp.org