2.7.1 Colocation & Clock Synchronization from Zero

Basic Prerequisites: None - this page assumes zero prior knowledge.

---

1. Intuition & Practical Objective

The single most important fact in all of high-frequency trading is a physics constant: information travels at the speed of light, $c=2.998\times 10^8$ m/s ≈ 300,000 km/sec ≈ 300 m per microsecond. Everything else in this folder is a consequence of that one number.

Three intuitions, each built on the previous:

1. Speed is distance. A signal covers 300 m in one microsecond. The moment you are physically farther than a rival from the exchange’s matching engine, you are literally slower - no amount of software fixes it. That is the entire reason “colocation” (renting a rack inside the exchange data centre) exists: at the microsecond frontier, being a few kilometres off-site is a death sentence, because 1 km of fiber costs you ≈5 µs.

2. Some information races are decided by milliseconds. The canonical example (Budish–Cramton–Shim 2015): the E-mini S&P 500 futures contract (ES) trades in Chicago, the SPY ETF that tracks the same index trades in New York. When the ES price jumps, the SPY price “should” follow instantly - they are near-perfect substitutes. But it takes light ~4 ms to cross the ~1,180 km between the two cities. During those 4 ms, one price is stale relative to the other, and whoever physically gets there first (or sees the jump first) captures a risk-free arbitrage. The prize isn’t intelligence - it’s being first.

3. To know who was first, clocks must agree. Suppose two venues timestamp events with their own clocks. If venue A’s clock lags venue B’s by 5 µs, then a true A-then-B sequence of 3 µs appears as “B first.” Clock synchronization (PTP, GPS, nanosecond timestamping) is the practice of forcing all the clocks that feed a consolidated view to agree within the µs/ns budget - otherwise you cannot reconstruct who actually moved first, which is a regulatory and forensics disaster.

The one-sentence essence. “High-frequency trading is a game where the winning move is having your electrons arrive a few microseconds earlier than your rival’s - so physical distance (colocation), physical medium (speed of light in air vs fiber), and clock truth (synchronization) determine the outcome far more than any algorithm.”

---

2. Mathematical Ground Truth & Derivations

Units - the mental model. Give yourself these anchors:

• $c=2.99792458\times 10^8$ m/s.
• In 1 ns (one billionth of a second): $0.30$ m - about one foot. A modern FPGA matching decision (~tens of ns) is a few metres of light.
• In 1 µs: $300$ m. The kernel-bypass/FPGA “speed-of-light room” - about three football fields.
• In 1 ms: $300$ km. Cross-city and cross-continental distances.

Why fiber ≠ light. In glass the speed is the phase velocity $v_{\text{fiber}}\approx c/n$ with refractive index $n\approx 1.5$, i.e. $v\approx 0.67c$. Microwave links in air travel at $v\approx 0.9997c$. Convert to per-kilometre latency:

$$v_{\text{fiber}}\approx 0.67c\Rightarrow 5.00\mu \text{s/km},v_{\text{air}}\approx c\Rightarrow 3.34\mu \text{s/km}.$$

The reach of an edge. If you have a $\Delta t$ head start, you can be $\Delta t/(\mu \text{s per km})$ kilometres behind your rival and still win. At a 1 µs edge: $\approx 0.30$ km (air) - i.e. even a microwave rival only ~300 m behind you forces a dead heat. At a 10 µs edge: ~3.0 km (air).

The latent race in equations. For a cross-venue arbitrage between markets $M_1$ (signal origin) and $M_2$ (the price that lags), the full capture time is

$$T_{\text{capture}}=T_{\text{prop}}(M_1\to M_2)+T_{\text{T2T}}(M_2)=\underset{\text{light transport}}{\underbrace{\frac{d_{12}}{v}}}+\underset{\text{orders to reach M2’s engine}}{\underbrace{(T_{\text{prop}}+T_{\text{nw}}+T_{\text{mach}})}}.$$

Only the first term is irreducibly about where the markets are; everything else is optimized by colocation and low-latency infrastructure from Low-Latency Systems Architecture.

---

3. Computational Implementation - units, reach & the “where is the race” question

Runs on the standard library only. This is the from-zero way to see the physics before any trading model.

The takeaway that makes colocation non-negotiable: a 1 µs edge is worth only ~300 metres of positional advantage. Since racks inside the building vs a data centre 10 km away is a 50 µs gap (fiber), nobody serious runs a latency arbitrage from off-site.

---

4. Failure Modes & First-Principles Breakdowns

1. Conflating “fast software” with “fast physics.” A brilliant algorithm on a machine 400 km away loses to a dumb one colocated, because the 1.3 ms propagation term dwarfs any software improvement. First-principles check: always split $T_{\text{prop}}$ out of the budget before blaming code.
2. Mixing up ns, µs, ms. A $10^{-3}$ slip converts a sub-µs edge into a milliseconds handicap. Unit discipline is the whole ballgame at this tier.
3. Assuming fiber == light. Fiber is ~$0.67c$, not $c$; the naive $d/c$ estimate understates real fiber latency by 50%. Use the per-km figures from §2.
4. Believing “whoever is faster wins.” If your clock says you won but the venue’s dueling timestamped clocks disagree, “fast” is meaningless - that is the subject of 04 · Clock Synchronization.

---

5. References

• Budish, Cramton & Shim (2015)
• O’Hara, Maureen (2015) - “High-Frequency Market Microstructure”
• MacKenzie, Donald (2021), Trading at the Speed of Light, ch 1–2

---

6. Connected Graph Bridges

• Continue: 02 · The Latency Race · 03 · Colocation & Networks · Index Hub
• Sibling system tier: Low-Latency Systems Architecture · Hardware Acceleration & FPGA
• Why microseconds matter economically: The Latency Race