The clustering feature which gives a score to remote clock sources based on their time offsets currently has one implementation, the 'discriminator' mode. The 'discriminator' mode filters out sources which are significantly off from a source which is trusted to provide a good time, even if it is not provided to a very high accuracy, e.g. NTP.
Where there are three or more high quality (e.g. PTP or PPS) sources available then a preferable form of clustering may be a 'voting' algorithm allowing the sources which diverge significantly from the majority of sources to be filtered out.
The task here is to add a 'voting' type of clustering which can be configured to this end. As with the discriminator, which is binary, there is likely to be some quantization needed in this determination so that the other selection criteria continue to be used effectively.
Implementation of this feature would improve the 'Enterprise Profile' offering in sfptpd.
The clustering calculation is currently performed on behalf of the sync instances by a callback in sfptpd_engine.c and defined in the global configuration sfptpd_general_config.c. It is possible that a future change will simplify these mechanics to avoid the callback and determine the score solely in the engine but this should not affect the implementation of this proposed enhancement.
The clustering feature which gives a score to remote clock sources based on their time offsets currently has one implementation, the 'discriminator' mode. The 'discriminator' mode filters out sources which are significantly off from a source which is trusted to provide a good time, even if it is not provided to a very high accuracy, e.g. NTP.
Where there are three or more high quality (e.g. PTP or PPS) sources available then a preferable form of clustering may be a 'voting' algorithm allowing the sources which diverge significantly from the majority of sources to be filtered out.
The task here is to add a 'voting' type of clustering which can be configured to this end. As with the discriminator, which is binary, there is likely to be some quantization needed in this determination so that the other selection criteria continue to be used effectively.
Implementation of this feature would improve the 'Enterprise Profile' offering in sfptpd.
The clustering calculation is currently performed on behalf of the sync instances by a callback in
sfptpd_engine.cand defined in the global configurationsfptpd_general_config.c. It is possible that a future change will simplify these mechanics to avoid the callback and determine the score solely in the engine but this should not affect the implementation of this proposed enhancement.