A strong answer. “Before optimising anything I'd want to know whether the gap is even ours to close, because a four-microsecond difference is often not in the code at all.

First, measure properly. Hardware timestamps at the NIC, in and out, not software timestamps in the application — software clocks are the same order of magnitude as the thing we're measuring. And I want the full distribution, not the mean. A p50 of 6 µs with a p99.9 of 40 µs is a completely different problem from a flat 6, and the second is usually what costs money, because the market moves precisely when everything is busy.

Then decompose the path. Wire to NIC, NIC to userspace, parse, decide, encode, send, NIC to wire. Each has its own budget. In practice the distribution is usually surprising: parsing and encoding often dominate, and the strategy logic — the part everyone assumes is the problem — is a few hundred nanoseconds.

Then the non-code causes, which I'd check before touching the code. Are we in the same colocation hall as them, on the same cross-connect topology? Are we taking the same feed — a consolidated feed versus a direct exchange multicast feed is microseconds of difference before any code runs. Is our switch adding a store-and-forward hop where cut-through would do? Is the NIC doing interrupt coalescing? Is the CPU scaling frequency, or is a core shared, or is the process migrating between NUMA nodes? Is there a page fault on a cold path? Half the latency problems I've seen were configuration, not code.

Then the code. No allocation in the hot path, no locks, no system calls, no branches you can hoist, and the data laid out so the working set stays in L1. Busy-poll rather than block, pin threads to isolated cores, disable the scheduler on those cores.

Then hardware. FPGA for feed parsing and for the simplest risk checks is real and well proven; FPGA for actual strategy logic is much rarer than the marketing suggests and only pays when the logic is genuinely simple.

And I'd ask the commercial question: is 2 µs the winning number, or is this a race we are not going to win? Sometimes the right answer is to compete on a strategy where four microseconds does not decide the outcome, and being willing to say that is not a weakness.”