Time Synchronisation of perfSONAR hosts

[ChristopherWalker]

To my mind, the current limiting factor for one way latency measurement is clock drift.

Things that could be done are:

Short term (ie for 5.3), Documentation on:

• Best practice chrony config
  • We've seen a factor of 4 precision by making some small changes.
• That PTP can be used, and pointers on how to set PTP up:

• Generic instructions (which could be RHEL manuals)
• Any perfSONAR specific config issues

Long term (and almost certainly not for 5.3)

• Hardware timestamp support to gain a bit more precision in OWAMP packets
  I think you've looked at this and decided it's non trivial
• A potential idea is that PTP transparent clocks report the residence time of a PTP packet - at least that's what I understand from PTP Feature Overview and Configuration Guide (p12). Assuming that's the case, I wonder if there's value in measuring that (or whether that's so niche or not supported it's not worth bothering).

[ChristopherWalker]

Raul made the following changes to the chrony config on our Slough and London servers:

server ntp5.ja.net  minpoll 5 maxpoll 5 iburst
server ntp6.ja.net  minpoll 5 maxpoll 5 iburst
# Previously minpoll/maxpoll not specified, so used the default maxpoll 6
# This doubles the frequency of polling the NTP server

# Lock chrony into memory
lock_all

Raul also changed (nothing to do with chrony):

 vm.swappiness from 60 to 1

We haven't established which of the changes made the difference.