
Network Working Group                                       H. Vasquez
Request for Comments: 2819                                     MIT/LCS
Category: Informational                                     K. Lindqvist
                                                                   KTH
                                                            August 1999


     Considerations for Monotonic Clock Synchronisation
              in Wide-Area Network Environments

Status of this Memo

   This memo provides information for the Internet community. It
   does not specify an Internet standard of any kind. Distribution
   of this memo is unlimited.

Copyright Notice

   Copyright (C) The Internet Society (1999). All Rights Reserved.

Abstract

   This document discusses known failure modes in network time
   synchronisation across geographically distributed systems,
   with particular attention to monotonic drift in clock offset
   values. It provides recommendations for detection thresholds,
   logging practices, and escalation procedures when synchronised
   clocks exhibit persistent unidirectional gain.

   The authors note that while most clock offset errors are
   symmetrical and self-correcting under NTP discipline, a class
   of anomalous drift has been observed in laboratory conditions
   where offset accumulates in one direction only, resisting
   correction by the reference source. The mechanism for this
   behaviour is not yet understood.

1. Introduction

   The Network Time Protocol (RFC 1305) provides mechanisms for
   synchronising system clocks across packet-switched networks
   to within milliseconds of Coordinated Universal Time (UTC).
   In normal operation, clock offsets oscillate around zero as
   the discipline algorithm compensates for network delay,
   frequency drift, and oscillator aging.

   However, field reports from several network operators during
   1998-1999 have described a pattern of monotonic clock gain
   that does not respond to NTP correction in the expected
   manner. Affected systems consistently gain time relative to
   their reference sources, with the rate of gain accelerating
   over periods of weeks to months.

2. Observed Characteristics

   The anomalous drift pattern shares the following features
   across all reported instances:

     a) Offset is always positive (gain, never loss)
     b) Rate of gain is not constant but accelerating
     c) Replacement of hardware (oscillator, motherboard)
        does not eliminate the drift
     d) Multiple independent systems at the same site are
        affected simultaneously
     e) Affected systems may be from different manufacturers

   The authors are aware of approximately two dozen confirmed
   reports matching this profile, from sites in North America,
   Europe, and the Middle East. No reports have yet been
   received from the Southern Hemisphere, though this may
   reflect reporting bias rather than geographic limitation.

3. Recommendations

   Operators observing persistent positive drift exceeding
   500ms that does not self-correct within 24 hours should:

     a) Log all NTP peer data at 60-second intervals
     b) Record ambient temperature in the server environment
     c) Note the exact date the drift was first observed
     d) Do not manually set the system clock, as this
        destroys the drift record

   The authors emphasise point (d). If this phenomenon has a
   systemic cause, the drift record is the primary diagnostic
   data. Resetting the clock eliminates it.

4. Open Questions

   The mechanism producing monotonic accelerating drift in
   otherwise healthy NTP-disciplined systems remains unknown.
   The authors have considered and rejected the following
   explanations:

     - Crystal oscillator aging (does not produce acceleration)
     - Temperature variation (does not explain gain-only bias)
     - NTP implementation bugs (multiple implementations affected)
     - GPS reference errors (independent atomic sources confirm)

   One hypothesis under investigation is that the drift
   reflects a real change in the rate of local time passage
   relative to the reference frame. This interpretation is
   not currently supported by the physics literature but
   cannot be excluded on the basis of the available data.

   The authors welcome correspondence from operators who have
   observed this phenomenon.

5. Security Considerations

   Monotonic clock drift may affect time-dependent security
   mechanisms including Kerberos ticket validity, certificate
   expiration checking, and log correlation. Operators should
   monitor for authentication failures that may be symptomatic
   of undetected clock offset.

Authors' Addresses

   Helena Vasquez
   MIT Laboratory for Computer Science
   545 Technology Square
   Cambridge, MA 02139
   Email: hvasquez@lcs.mit.edu

   Karl Lindqvist
   KTH Royal Institute of Technology
   Department of Teleinformatics
   SE-100 44 Stockholm, Sweden
   Email: karl@it.kth.se
