A colocated strategy is 200 nanoseconds slower than a competitor and both are in the same data centre. List everything that can still differ between two colocated participants, in order of how many nanoseconds each is worth.

A colocated strategy is 200 nanoseconds slower than a competitor and both are in the same data centre. List everything that can still differ between two colocated participants, in order of how many nanoseconds each is worth.

Approach: Walk the path from the exchange's publication point to the order reaching the matching engine, and note every place two participants can be treated differently or build differently.

The largest differences are the implementation technology and the switch path, worth hundreds of nanoseconds each, while cable length is worth only a few. An FPGA that parses the feed and emits an order without a general purpose processor completes tick to trade in 100 to 300 ns, where a well tuned software path with kernel bypass is 1 to 5 microseconds, so that single choice dwarfs everything else. Next is the number of switch hops: a cut-through switch adds 300 to 500 ns per hop and a store-and-forward switch adds the serialisation time of the whole frame, so a direct cross connect to the exchange handoff rather than a path through two aggregation layers saves a microsecond. Then the NIC itself, where models differ by several hundred nanoseconds in receive to application delivery. Then the feed choice, since a full depth feed carries more bytes to parse than a top of book feed. Then protocol encoding, where a binary order format is faster to build than a text one. Cable length is last: light in fibre covers about 5 ns per metre, so a 20 metre difference inside a hall is 100 ns one way, which is why exchanges equalise cable lengths within a colocation room. Ordering the list this way is the point, because teams often optimise the cheapest term first.

Follow-up: Your competitor reacts to a trade print in 90 nanoseconds, which is less than your NIC's receive path. What must they be doing?

Key concepts: colocation, cross connect, fpga, cut through.