Lakshya

Trading & Capital Markets · Chapter 8 of 8

Latency: where it comes from, and what it costs to remove

A budget in microseconds, an arms race with a physical floor, and the honest question of whether you should be in the race at all.

6 min read1 diagramAll 8 chapters

Latency is the most romanticised subject in trading and one of the most misunderstood. It matters enormously for a small number of strategies and is close to irrelevant for most others. Being able to say which is which is more valuable than knowing the numbers.

TICK-TO-TRADE, DECOMPOSED — THE NUMBERS ARE ORDERS OF MAGNITUDE, NOT QUOTES retail broker app → exchange 50–300 ms internet, broker queue, RMS colocated, ordinary stack 200–800 µs kernel networking, no tuning colocated + kernel bypass 5–20 µs Onload / DPDK, busy-poll, pinned cores FPGA feed parse + risk ~1–3 µs hardware in the path the exchange's own matching sub-µs internal you cannot go below this CHICAGO ↔ NORTHERN NEW JERSEY — WHERE PHYSICS BECAME A BUSINESS ordinary fibre~14–16 ms round tripfollows railways and roadsstraight-line fibre~13 msa cable dug in a straight linemicrowave / millimetre~8–9 msair is faster than glass; rain fades it Light in glass travels at about two-thirds of c. Microwave through air is close to c — which is the entire reason towers were builtacross Ohio.
A latency budget, and the Chicago–New Jersey route where physics became a business. The figures are orders of magnitude rather than quotes. Note the bottom row: the exchange's own matching time is a floor you cannot go below.

Why time matters at all

Because of price-time priority. At a given price, the queue is ordered by arrival. Being faster does not get you a better price — it gets you ahead of everyone else who wants the same price. Two consequences follow that people frequently miss:

  • Latency matters most for strategies that compete for the same fill. Market making, arbitrage between correlated instruments, and reacting to a public event. If your edge is that you know something others do not for the next three days, microseconds are irrelevant.
  • It is relative, not absolute. Being 5 microseconds slower than the fastest participant in a race you are competing in is fatal; being 50 milliseconds slower than them in a strategy where you are not competing costs you nothing.
  • Queue position is the actual currency. A market maker cares about being early in the queue at a price level, because the queue is served in order and the last person in it gets filled precisely when the price is about to move against them.

Where the microseconds actually are

The distribution surprises people. The strategy logic — the part everyone assumes is the bottleneck — is usually a few hundred nanoseconds. The time is elsewhere:

  • Wire to application. The kernel network stack costs several microseconds and a context switch. Kernel bypass — Solarflare Onload, DPDK — removes most of it by mapping the NIC directly into the application's address space.
  • Parsing. Decoding the exchange's binary protocol. Often the largest software component, which is why it is the first thing moved to an FPGA.
  • Encoding and sending. The same in reverse.
  • Jitter, which matters more than the mean. A garbage collection pause, a page fault, a CPU frequency change, a thread migrating across NUMA nodes, an interrupt. These produce a long tail, and the tail happens exactly when the market is busiest — which is when it costs you. Anyone who quotes a mean latency without a p99.9 has not measured properly.

The physical layer, where the money really goes

  • Colocation. Your servers in the exchange's own data centre. NSE and BSE both offer it in India; in the US the relevant halls are the exchange data centres in New Jersey and the CME's in Chicago. Exchanges equalise cable lengths within the hall precisely so that rack position confers no advantage — a regulatory-driven fairness measure.
  • Cross-connects. The physical cable from your rack to the exchange's access switch. Priced monthly, and the count adds up.
  • Direct feeds versus consolidated. In the US, taking each exchange's own feed and building your own view is faster than the consolidated tape. This is a data cost, not an engineering one, and it is the largest single latency decision most firms make.
  • Inter-city links. The Chicago to New Jersey route is the famous one. Straight-line fibre beat the railway-following route; microwave beat fibre because radio through air travels close to the speed of light while light in glass travels at about two-thirds of it. Microwave carries far less capacity and degrades in heavy rain, so firms run both and fail over. Hollow-core fibre is the more recent development, closing part of the gap without the weather problem.

The honest question, and the one that reads as senior

Every latency conversation should include: is this a race we can win, and is winning it worth what it costs?

The arms race has a physical floor. Once you are colocated, on direct feeds, with kernel bypass and FPGA parsing, the remaining differences are small and the cost of each further microsecond rises steeply. A firm entering late is buying into the expensive end of a curve against incumbents who bought the cheap end years ago.

The strategically interesting answer is frequently to compete somewhere else — on a signal with a longer horizon, in a less contested instrument, or on execution quality rather than speed. Being willing to say that in an interview at a firm that does race is not a weakness; it demonstrates that you understand what the spend buys.

Measuring it properly

  • Hardware timestamps at the NIC, in and out. Software timestamps are the same order of magnitude as the thing being measured.
  • PTP for clock synchronisation, not NTP. NTP is accurate to milliseconds; you need sub-microsecond, and unsynchronised clocks make cross-machine measurements meaningless.
  • Histograms, always. Report p50, p99, p99.9 and max. A mean hides exactly the behaviour that costs money.
  • Measure in production, continuously. Latency regresses silently as code and configuration change, and a benchmark run once is a number about the past.

Check yourself

Why does latency matter, mechanically?Price-time priority. At a given price, arrival order decides who gets filled, so speed buys queue position rather than a better price.
Why is jitter worse than mean latency?Because the tail occurs when the market is busiest, which is exactly when the fill is worth most. A p99.9 of 40µs on a 6µs median is a much worse system than a flat 10µs.
Why microwave between Chicago and New Jersey?Radio through air travels near the speed of light; light in fibre travels at about two-thirds of it. Lower capacity and rain fade are the trade-offs, so firms run both.
When does latency not matter?When your edge has a horizon of hours or days and you are not competing for the same fill. Most systematic strategies are in this category, and spending on speed there is misallocated.
← Market data: what it is, what it costs, and who sells itPractice bank →