The limits of NTP accuracy on Linux
Summary
Scott Laird examines the practical limits of high-precision NTP synchronisation on ordinary Linux systems. NTP uses four timestamps to estimate offset and delay. System firmware can cause unobservable pauses of several hundred microseconds.
Ideas
- NTP can only estimate network delays under assumptions about the outbound and return paths.
- Asymmetric paths cause systematic time errors despite stable measurements.
- Hardware timestamps avoid variable delays in the software stack.
- SMBIOS activity can halt the visible computer for hundreds of microseconds.
- Inexpensive systems need measurements because platform latencies are hardly predictable.
Insights
- A precise source does not guarantee a precise clock on the end system.
- Time accuracy is a property of the entire measurement chain.
- Stable deviations can be more dangerous than clearly visible jitter.
Facts
- PTP and hardware timestamps address different accuracy classes from ordinary NTP.
Recommendations
- First define the accuracy and traceability you actually need.
- Test platform, network path and time source together under realistic load.
References
Links to the original source and the Web Archive open in a new tab.