What are virtual machines?

What is an effective means for evaluating time drift across thousands of virtual machines?

• I would like to evaluate time spread between a large number virtual machines. Can I find this information within NTPd or is there a way to collect time drift on the client itself? I would like to emphasize that the machines in question are virtual machines, so they don't have direct access to the clock.

• Answer:

  The common implementation of the Network Time Protocol (NTP) can provide this information. NTP believes in absolute time, and attempts to carefully measure and control the clock frequency & drift of a given computer against an absolute notion of the correct (UTC) time. You can collect the system clock drift information with ntpdc(1) remotely over the network, or ntptime(1) on a given local system. NTP time synchronization accuracy (and thus drift measurement) is necessarily limited by both: (most pee cees have terrible system clocks, driven by cheap, uncalibrated, uncompensated crystals), and add further noise by poor interrupt processing, bad process scheduling, and/or other implementation flaws (ignorant, careless ... I mean "systems programmers"). Putting a in the middle of that is most likely to make things yet worse. NTP will accurately measure just how bad all of this is in total, but it cannot tell you the source of the drift: Is it a cheap crystal whose frequency wanders all over the place? Is it noise/drift induced by bad OS implementation? A combination of both? All NTP can do is measure the effect(s), and report. Just to complicate things yet more, you can have bias and jitter in the reference time samples introduced by the network interface cards and device drivers (and network protocol stack) that NTP uses to talk to the time references outside the local system (your NTP "peers"). NTP compensates for bias from network congestion by taking lots of samples over a relatively long baseline. It's quite amazing how subtle time & frequency measurement can be. "... chasing the decimals down into the noise." - Dr. Dave Mills,