Nearly every broker or prop firm setup has a mix of timestamps from application servers, network devices, gateways and external venues. Those clocks may all display UTC, yet differ by milliseconds or record events at different points in the processing path. A consistent time source is only part of the solution. Operators also need to know how a timestamp was captured, how uncertain it is and how to compare it with timestamps from another system.
The right design depends on the job. NTP is often sufficient for business systems and general logging. PTP can provide tighter synchronisation for market data and execution paths. GNSS can supply a trusted reference, but it does not remove distribution, monitoring or holdover requirements. A useful comparison starts by separating clock accuracy from timestamp quality.
Separate clock accuracy from timestamp meaning
Clock accuracy describes how closely a host's time matches a reference. Timestamp meaning describes the event being marked. A gateway might record a packet when its network interface receives the final bit, while an application records it after parsing and queueing. Both timestamps can be accurate to the same UTC reference and still differ by tens or hundreds of microseconds because they describe separate events.
Wall-clock time is used for calendar timestamps and cross-system correlation. A monotonic clock measures elapsed time and should be used for durations, timeouts and latency histograms. Wall clocks can step after a correction or leap-second handling. A monotonic clock should not move backwards, though its origin is normally local to the host and cannot be directly compared across machines.
For each timestamp field, document at least the clock source, capture point, resolution and uncertainty. Also record whether the value is UTC, a venue's time scale, or a local monotonic counter. An event schema that stores only an integer timestamp without these details can make an apparently precise audit trail misleading.
Time corrections need careful handling. NTP clients commonly slew the clock, adjusting its rate gradually, but may step it during initial synchronisation or after a large error. PTP clients can discipline the system clock or expose a hardware clock to applications.
Record offset and synchronisation state alongside events where possible. If a system changes time source or loses lock, that transition should be visible to operations and downstream consumers.
Compare the available synchronisation approaches
The figures below are typical market ranges, not service guarantees. Actual results depend on network path asymmetry, hardware, configuration, operating-system behaviour and measurement point. Offset means the host's estimated difference from its reference, not end-to-end application latency.
| Approach | Typical offset range | Typical deployment burden | Common limitation |
|---|---|---|---|
| Internet NTP | 1 to 20 ms | Low, usually minutes to hours | WAN path variation and unknown asymmetry |
| LAN NTP | 0.1 to 1 ms | Low to moderate, hours to a few days | Software scheduling and network congestion |
| Software PTP | 10 to 100 µs | Moderate, days to weeks | Host scheduling and switch behaviour |
| Hardware PTP | 100 ns to 1 µs | Moderate to high, days to weeks | Profile, NIC and switch compatibility |
| GNSS-fed grandmaster | 100 ns to 1 µs at a disciplined host | High, including antenna and resilience work | Antenna failure, interference or loss of satellite view |
NTP is the practical default for systems where millisecond-level alignment is adequate. A pair of internal NTP servers, fed from separate upstream sources, avoids making every host dependent on a public service.
Chrony is commonly used on Linux and can expose offset, root dispersion and source state. Do not treat a small reported offset as proof of low uncertainty: the reference path and its quality matter.
PTP, defined by IEEE 1588, is suited to tighter requirements. A grandmaster supplies the reference, while boundary clocks or transparent clocks in switches can limit the effect of network residence time. Hardware timestamping at the NIC reduces variation from software scheduling. Typical deployments use a controlled network and a consistent PTP profile. Mixing profiles, relying on unmanaged switches or allowing asymmetric routes can erase much of the benefit.
GNSS is a reference source, not a complete timing architecture. A roof antenna and receiver can discipline a grandmaster, which then distributes time over PTP. A resilient installation usually has more than one reference path, monitored antenna status and a defined holdover plan. Holdover is the period a clock can remain within its required tolerance after losing its reference. Its quality depends on the oscillator, temperature and time since the last good reference.
Some infrastructure teams combine approaches: GNSS-fed PTP for market data capture and execution hosts, with NTP for monitoring, office systems and less time-sensitive services. The platform and its operational constraints matter as much as the protocol. For broader coverage of broker infrastructure design, see Trading Technology's infrastructure reporting.
Reconcile timestamps across the event path
A trading event may pass through a feed handler, normaliser, strategy, order gateway, broker network and venue. Each system can add a timestamp, but those values are not automatically comparable. The venue's acknowledgement time might use a different clock source, time scale or definition of receipt from the broker's gateway timestamp.
Keep raw timestamps immutable. For every source, maintain a mapping to a common timeline and store the correction separately.
A record can carry the original timestamp, source identifier, estimated offset and uncertainty. This allows an investigation to reproduce the conversion used at the time rather than silently rewriting historical data when a clock model changes.
For example, suppose a gateway reports an outbound event at 10:00:00.100200 UTC with an estimated offset of +40 µs, while a venue reports receipt at 10:00:00.100500 with an uncertainty of 80 µs. After applying the gateway correction, the apparent transit time is about 340 µs.
The combined uncertainty is large enough that the result should be reported as an estimate, not as an exact one-way latency. If clock errors are correlated, calculate that explicitly rather than assuming independent errors.
Useful reconciliation controls include:
- Define event semantics for every timestamp: packet arrival, decode completion, strategy decision, socket write, NIC transmit or venue receipt.
- Keep UTC wall-clock timestamps for cross-system correlation and monotonic counters for local durations.
- Store source clock, offset estimate and uncertainty with captured events or in a linked time-quality record.
- Preserve exchange-provided timestamps unchanged, alongside the venue, field definition and time scale.
- Flag negative durations, impossible event ordering and offsets outside the system's tolerance. Do not silently clamp them to zero.
Timestamp reconciliation should not be used to infer causal order where the uncertainty intervals overlap. If event A is stamped 20 µs before event B but their combined uncertainty is 50 µs, the data does not establish which happened first. Use sequence numbers, packet order or a shared hardware capture point where the question requires stronger evidence.
Build monitoring and recovery into the design
Deploy timing in stages. Establish a tolerance for each workload, then test both steady state and failure conditions. A strategy host that needs microsecond-level event ordering has a different requirement from a reporting database that only needs reliable minute-level correlation.
- Set requirements by service. Define maximum offset, uncertainty and recovery time for market data, order entry, audit logs, monitoring and back-office systems. State whether the limit applies to the host clock or the timestamp at a particular capture point.
- Choose sources and topology. Use independent upstream references where appropriate. For PTP, document the grandmaster selection, profile, domain, network path and which devices provide hardware timestamping. Keep a lower-precision fallback only if its behaviour is understood.
- Validate the measurement path. Compare the host's reported offset with an independent measurement, such as a timing analyser or a second disciplined capture system. Check for asymmetric paths, NIC timestamp configuration errors and switch ports that do not support the intended PTP behaviour.
- Exercise failure modes. Disconnect the GNSS antenna, stop an upstream NTP source, change the active grandmaster and test server reboot behaviour. Confirm that the clock enters holdover or raises an alert, rather than appearing healthy while its uncertainty grows.
- Monitor quality, not just reachability. Track offset, dispersion or estimated uncertainty, source changes, frequency correction, PTP state and time since last reference. Alert before the service limit is crossed, and include the relevant host and network details in the alert.
- Check the application record. Verify that application and NIC timestamps use the expected clock, that timestamps are not overwritten during normalisation, and that event ordering survives ingestion, replication and replay.
Typical operations policies set warning thresholds below the maximum permitted offset, with tighter limits for execution and market data than for general logging. The precise threshold is workload-specific. A stable offset of 20 µs may be acceptable for one service and a fault for another. Review the error budget against measured uncertainty, not just the nominal precision advertised for a timing device.
Keep a record of configuration changes, source failovers and clock alarms with incident data. When a latency report conflicts with venue or network evidence, first compare event definitions and uncertainty bounds. A clock can be synchronised and a timestamp still be unsuitable for the question being asked.
This advice does not apply unchanged to a venue-specific or regulated environment that mandates a particular timing source, PTP profile, traceability standard or timestamp method. In that case, the venue specification and applicable rules take precedence over a general-purpose design.
Frequently Asked Questions
What is the difference between NTP and PTP?
NTP is simpler to deploy and is typically adequate for millisecond-level alignment. PTP can provide tighter synchronisation, particularly with hardware timestamping and compatible network switches, but needs more controlled configuration and monitoring.
Is GNSS enough to synchronise trading servers?
No. GNSS can provide a reference to a receiver or grandmaster, but time still has to be distributed to servers and monitored. Antenna failure, interference and loss of satellite view also make holdover and an independent reference path important.
Should trading systems use UTC or monotonic time?
Use UTC wall-clock time to correlate events across systems. Use a monotonic clock to measure local durations and timeouts because wall-clock time can be adjusted. Record enough metadata to identify the clock and capture point behind each timestamp.
How should exchange timestamps be reconciled with broker timestamps?
Keep the original venue timestamp and document its time scale and event definition. Apply a separately recorded offset model to local timestamps, including an uncertainty estimate. Avoid claiming a precise event order if the uncertainty ranges overlap.
Platforms · Software · Providers · Markets