Notes on "Latency Limbo: How Low Can You Go?" @ FPL Americas 2009

Lasalletech had a great time at the FPL Americas 2009 conference. While we were pretty busy showing off our new product suite, I did manage to sit in on a couple sessions. Here are some interesting points from the panel on Low Latency trading:

There are two aspects of latency that you can measure
1) How fast are you getting prices
2) How fast can you execute

However, the critical measurement of how successful you are is whether your fill rates are good. Latency is relative, not absolute. It's all about how fast you are compared to your competition. As a matter of fact, the relative ranking is so important that Eurex will start distributing anonymous standings of where you rank against others on their exchange.

All the speakers reiterated the basic truths that all good engineers should already know.

Do not optimize without understanding where your bottlenecks are. It makes no sense to optimize your network only to find out that your algorithm is where the traffic jam is.

Always measure latency under load. Often, the numbers look great when there is no stress on the system

Latency is a statistical distribution. Do not just look at the mean latency. You need to understand the distribution curve. As a matter of fact, you can use some basic quantitative analysis to predict latency, given enough data. Latency tick databases are becoming more common.

Finally, one of the panelists made an interesting point - today most of the exchanges out there use price/time priority for matching. The exchanges are now forced to upgrade their infrastructure because of the latency arms race. However, it is probably the case that this costs them more money than it is worth. It is not inconceivable that the exchanges can switch to a price/size priority and all these algorithms will likely fall on their face.