Time Synchronisation in Distributed Trading Systems

Clocks on different trading hosts drift by microseconds or more, which makes a raw cross-host timestamp comparison meaningless unless the fleet is actively synchronised.

Every host in a trading system keeps its own clock, and those clocks drift apart from each other by microseconds to milliseconds over the course of a trading day unless something actively pulls them back into alignment. Comparing a timestamp written on the feed handler host to one written on the risk host, without knowing how far apart their clocks actually are, produces an ordering of events that can be simply wrong, which is why time synchronisation is treated as core infrastructure rather than an operational afterthought.

A server's local clock is driven by a crystal oscillator that runs slightly fast or slightly slow compared to true time, and that error compounds continuously between corrections. Two hosts that were synchronised to within a microsecond at nine in the morning can be tens of microseconds apart by the close without a correction mechanism running in between, and for a system making decisions in single-digit microseconds, tens of microseconds of clock error is larger than the decision itself.

This matters most when reconstructing the order of events across machines: which arrived first, a market data update on the feed handler or a risk check on a separate host. Without synchronized clocks the answer depends on whose clock happened to be further ahead that second, not on what actually happened.

Network Time Protocol keeps clocks synchronized to within roughly a millisecond over a typical network path, which is adequate for logging and general operations but too coarse for the microsecond-level comparisons a trading system needs. Precision Time Protocol, defined in IEEE 1588, is designed for local networks with hardware timestamping support in the network interface cards and switches, and a well-tuned PTP deployment inside a colocation facility can hold clocks to within a few hundred nanoseconds of each other.

The gap between NTP and PTP is not a rounding difference. A firm running latency-sensitive strategies typically runs PTP across the trading fleet and reserves NTP for hosts where microsecond precision is not required, since PTP hardware and the switches that support it cost more and need dedicated configuration.