Client-server clock synchronization: estimate time without hiding uncertainty

Map one clock onto another

A networked game needs a shared interpretation of event timestamps, but clients do not begin with identical clocks. Use a monotonic local clock for elapsed time and maintain a mapping to server session time. Calendar time is unnecessary for many match events. Clock offset describes a time difference; drift describes a rate difference. Synchronizing once does not eliminate later drift.

Keep the time domain explicit in data structures. A server timestamp, a local input timestamp and a delayed presentation timestamp are not interchangeable merely because all use seconds.

Estimate offset with a timestamp exchange

Let t1 be client send time, t2 server receive time, t3 server reply time and t4 client receive time. With clocks advancing at nearly the same rate during the exchange, the standard estimator is offset = ((t2 - t1) + (t3 - t4)) / 2, with server time minus client time as the sign convention. Network round-trip delay is delay = (t4 - t1) - (t3 - t2). Subtracting server processing time avoids treating that interval as network travel. These are the timestamp relations described by NTP. [1]

Worked example: processing delay and path asymmetry

Take illustrative timestamps of 10.000, 10.260, 10.280 and 10.120 seconds in the stated order. Server processing occupies 0.020 seconds. The estimated network round trip is 0.120 - 0.020 = 0.100 seconds, and the offset is (0.260 + 0.160) / 2 = 0.210 seconds.

The offset estimate assumes symmetric outward and return delays. Paths taking 70 milliseconds outward and 30 milliseconds back would bias that estimate by 20 milliseconds. The exchange alone cannot identify which direction contributed the delay. A small round trip reduces some uncertainty; it does not prove that the paths were symmetric.

Filter measurements and apply corrections deliberately

Collect repeated samples, match replies to outstanding requests and reject obsolete session data. Track delay variation and offset dispersion. Mirror exposes a synchronized double-precision time estimate, round-trip information and measurement variation, and documents smoothing with an exponential moving average. Preserve the precision of the supplied time type. [2]

Filter noisy samples rather than snapping the gameplay clock to every reply. Decide whether small corrections adjust the mapping gradually and how a large discontinuity resets dependent buffers. A monotonic gameplay timeline must not accidentally move backward when an offset estimate decreases. Resynchronize after suspension, reconnection or a server epoch change; invalidate comparisons across incompatible epochs.

Validate the consumers of time

Test clock drift, long sessions, server processing spikes and asymmetric links. Log raw samples, filtered estimates and corrections separately. Measurement variation is not a certified bound on true offset error. The server must still validate client-supplied event times. Synchronization provides a time estimate; it does not itself choose an interpolation delay, authorize historical actions or make a client's clock authoritative.

Sources

• [1] Network Time Protocol Version 4: Protocol and Algorithms Specification — David L. Mills, Jim Martin, Jack Burbank and William Kasch, IETF RFC 5905.
• [2] Time Synchronization — Mirror Networking contributors.