The Gateway to Algorithmic and Automated Trading

Latency: The Hands-on Approach

Published in Automated Trader Magazine Issue 05 April 2007

Obviously here at AT we get a lot of press releases along the lines of “XXX announces the launch of latency busting YYY data feed/network/event processing engine that can process 10 billion messages/data points per nanosecond”.

That set us thinking. In other industries - often those related to technology or cars - you find "enthusiasts". They don't just buy the latest and best technology; instead they spend ages tinkering with what they already have. The end result is usually a hot rod tractor that explodes, or a ludicrously overclocked processor that trips up the operating system every five seconds.

So, what if similar "enthusiasts" existed in the world of automated/algorithmic trading? If they did, they wouldn't be buying that YYY data feed/network/ event processing engine. Oh no, they would be round the back of the server racks with a soldering iron and a black cockerel.

Well we didn't have to look far before we found one. One of the joys of AT is our team of "secret squirrels" - insiders who are happy to give us the gossip but don't want the publicity (or in some cases, the broken legs that might come with it). One of our regular squirrels who runs a prop shop invited us round to their premises for the first time recently. As we stepped through the door, all our antennae were picking up a strong "enthusiast" signal because the place was a tip. Dismembered servers and discarded spanners littered every horizontal surface.

So we seized the opportunity to ask a few leading questions about how you might go about reducing trading latency if you were handy with a screwdriver. For obvious reasons we didn't get chapter and verse, but there were a few pointers. One of the most interesting was the general observation that much of the technology used for trading wasn't specifically for trading but was general purpose. Most network switches had a lot of functionality you didn't need, most operating systems by default ran background processes that were irrelevant, and network protocols had capabilities that if you ever needed them you could reproduce more efficiently another way.

As a result, rather than spending $Xm on the latest new latency busting technology you could get the same or better result by turning off irrelevant services in Windows (why would you need IIS on a trading workstation?), wrapping tinfoil round your network cables (well sort of) and hot wiring your router. We actually saw this last remedy in action. The gateway router to one exchange the shop trades on looked as if it was having a heart bypass. Extra pieces of cable had been soldered to the PCB, bypassing circuitry that was presumably deemed irrelevant and there appeared to be three network cables glued into the WAN port with builder's mastic.

To us it looked more like butchery than honing network performance, but since this shop makes shed loads each day and has been arbing everybody else off the pitch for the last five years, we kept our opinions to ourselves…