Hacker News Reader

Comment by leokun | Hacker News Reader

> While many UHFT firms have moved towards custom hardware (both for processing and networking), some less latency-sensitive systematic trading firms still make use of multithreaded C, C++ and Java (with custom garbage collection).

Aside from the hiring theme I have a question. I'd think given that network speeds would start being the bigger bottleneck. Is location at all important to these firms? The speed of light is fast, but the closer you are to the exchanges, the better the advantage over competitors must be.

View on HN

All the major exchanges offer co-location services, where you can put your trading engine (i.e. the "black box" that executed your trading strategies) in the same building as the exchange's trade matching engines. That means that you'll receive market data quicker and the exchange will receive your orders (or cancellations) quicker than someone who's located a few hundred miles away.

For more info, see: NYSE - http://nysetechnologies.nyx.com/en/infrastructure-solutions/...

London Stock Exchange - http://www.londonstockexchange.com/products-and-services/con...

Eurex - http://www.eurexchange.com/exchange-en/technology/co-locatio...

CME - http://www.cmegroup.com/globex/trading-cme-group-products/co...

Such co-location obviously works when you're talking about a single market but if you're looking to execute trading strategies across multiple markets, then you're going to be in the market for low-latency network connectivity between London and New York - http://www.telegraph.co.uk/technology/news/8753784/The-300m-... - or NYC and Chicago - http://gigaom.com/2012/02/10/wall-street-gains-an-edge-by-tr...

However, it doesn't really matter whether you shave 10ms off your network latency if your software is sluggish, so a lot of effort goes into optimising code and creating execution environments that minimise latency caused by abstraction layers. Companies use real-time operating systems and TCP/IP stacks that are implemented on hardware. They're now starting to use FPGAs and I'd be surprised if someone hasn't already splashed out on an ASIC.

I can't find the story now, but I read about an HFT making a play on the bid/ask spreads between new york and chicago, they spent a large sum of money to build their data center somewhere in the appalachians to save themselves something like 3 ms in total transit time.

I think you are referring to an article called "Barbarians at the Gateways" by Jacob Loveless, that recently popped up on HN:

- Web: http://queue.acm.org/detail.cfm?ref=rss&id=2536492

- PDF: http://portal.acm.org/ft_gateway.cfm?id=2536492&type=pdf

Distance between the exchanges is a factor, I believe there is a minimum distances each company can be between the exchanges. http://www.youtube.com/watch?v=aq1Ln1UCoEU This video at 23:27 talks about minimum 1000ft cable.

You're correct, co-location at the exchange data centre is indeed a requirement for most HFT firms.

I think (but I'm not 100% sure) that the length of network cabling is also tightly regulated at many exchanges so that nobody gains an advantage in that manner. If anybody has more insight into this, I'd love to hear about it.

That's correct. All the servers are connected using the same length cable so that the server at one corner won't receive the data a tiny bit faster than a server at the far corner. So there are simply plenty of unused cable next to the machines.

Generally it's possible to colocate servers in the same datacentre as the exchange servers. This isn't really a competitive advantage though, as anyone who needs latency that low will be doing the same, and the costs will be relatively insignificant compared to salaries of people who can build the required systems.

While I completely agree that there is no advantage over other HFT competitors provided by being located within the exchange datacentre, there is a disadvantage to not being there. This gives the exchange an opportunity to make a significant amount of money from each HFT company without hurting other customers.

There's a great Wired article about this from last year: http://www.wired.com/business/2012/08/ff_wallstreet_trading/...

They explain how firms built private fiber networks between Chicago and New York to shave a few milliseconds off latency.

From what I've read on the subject, yes, location is critical. As I understand it, HFTs are co-located in the exchanges' datacenters.

It is. There's a massive colo right next door to the NYSE.

The NYSE matching engines are in Mahwah, NJ, about 20 miles from the NYSE floor. If you rent colo space there's you're actually in the same building.

Reply on news.ycombinator.com