High-frequency trading has long stood as a great example of how the physicality of computing really matters, contrary to companies’ breezy declarations that everything is in The Cloud. Take for example this great anecdote from a long article on an impending class-action suit over HFT in The Guardian:
The NYSE was the first stock exchange to build its own data centre, at Mahwah, New Jersey (with an identical European sibling in Basildon, Essex). Finding engineers to describe what goes on inside this 37,000 sq metre (400,000 sq feet) hangar is not easy, but once you have, the eccentricity of the microverse comes comically alive. NYSE began by building a room called a pod, in the centre of which was the “matching engine”, a central server that gathers data and distributes it to market participants for a fee. For a higher fee, however, HFT traders could “co-locate” their own algorithmic computers inside the pod. And because space inside Pod 1 sold quickly, a second was built. When the traders in Pod 2 complained at being farther from the matching engine, a Pythonesque comedy began. “I always thought that was funny,” laughs one engineer down the phone, “because a foot of cable equates to one nanosecond’s difference – a billionth of a second. And when this first happened, the machines most of these guys used couldn’t even differentiate nanoseconds.”
But the NYSE centre is regulated, unlike private, third-party data centres, and has to be seen to be fair. So NYSE came up with an elegant solution, by measuring the distance to the farthest HFT server and giving everybody the same length of cable. This worked until engineers noticed that traders now wanted their servers located farther from the matching engine, because light signals bouncing around inside a long, coiled cable would in principle degrade and slow infinitesimally in comparison with a straight cable. In response, NYSE ran traders’ cables around the walls to minimise bend. Laugh at this Marx Brothers farce, however, and systems people will remind you that computers don’t work the way we do. Our “time” is meaningless to them: a nanosecond might as well be a second, or a century, because the message that arrives first gets the trade.