## Tuesday, December 11, 2012

### A Little Fact Checking...

This article caught my eye this morning.  It claims that there's a new market for point-to-point microwave links based on the fact that they're significantly faster than fiber optic connections.  By faster they mean latency (how long a message takes to get from sender to receiver), not bandwidth (how fast the content of a message takes to transmit).  The customer in this market is the high frequency trading world (aka “algorithmic trading”).

The need for low latency amongst high frequency traders makes perfect sense.  Once an electronic order arrives at an exchange, the actual trade can happen in less than a millisecond.  But does a microwave link actually have a significantly lower latency than a fiber optic link?

The article cites two factors that make microwave links faster:
1. The speed of light in air is about 300,000 km/sec, but in fiber is about 200,000 km/sec.
2. The route of microwave links tends to be more direct than fiber routes, which usually follow the right-of-way alongside railroads or highways.
The first reason is simple physics.  The second is true, but less certain.  There are also a couple of factors left out of this equation: both microwave links and fiber optic links require signal repeaters along long routes, but microwave links require them at smaller intervals – and each signal repeater introduces a small delay.

So let's do some math on the big market cited by the article: connections for traders in Chicago who are trading on a New York exchange.  The direct distance between Chicago and New York is about 1,150 kilometers.  Here's our math for each approach.

Fiber optic:
distance = 1,150 x 1.3 (routing length factor) = 1,495 km
latency = 1,495 / 200,000 = 7.45 ms
repeaters = 2 (every 500 km) @ 0.2 ms = 0.4 ms
total latency = 7.85 ms

Microwave
distance = 1,150 x 1.05 (routing length factor) = 1,207 km
latency = 1,207 / 300,000 = 4.03 ms
repeaters = 4 (every 250 km) @ 0.2 ms = 0.6 ms
total latency = 4.63 ms

Microwave beats fiber by 3.22 ms, one way - over three times the trading latency.  Yup, that's significant.

But there's one way to make this much, much lower latency, which the article doesn't mention at all – and it's the most common way by far: locate the high frequency trading servers physically nearby the exchange servers.  The electronic exchanges all offer (very expensive!) rack space in or near their datacenters for exactly this purpose...