Introduction

I was interested in the process behind building a bitcoin exchange that was actually fast (this was in the MtGox days) so I wrote a very simple single asset limit trading app in .NET for fun, and learned a bunch of interesting things on the way.

Market Background

There are only two order books:

• Buy
• Sell

There are many different kinds of order types in a real market but this app only does the basic limit orders. Adding in market orders would be pretty trivial as all you would have to do is use either Min() or Max() on the respective book. Another thing to look at would be to add the order constraints Fill or Kill and All or None. Right now, the orders are partially filled.

The Order Books

For the order books I used a SortedSet which in .NET uses a Red-Black Tree data structure. A SortedSet/Red-Black Tree is a self-balancing binary search tree, which basically means that as orders are added to the books they are automatically sorted for us. The orders aren't primitive types so for the Order class we needed to implement an IComparer that we aptly called ByLimit which is passed into the books on initialization.

Time

This was actually a huge surprise to me while building this application because I'm not used to developing under such tiny time constraints. My first thought was of course to use a DateTime, but found out that it is not precise at all (when dealing with time on a nanosecond level). In fact, DateTime's only go down to about a 16 millisecond "precision", which for trading is abysmal. This led me to investigate DateTime's, and the difference between precision and accuracy.

I ended up having to use StopWatch because the trades were executing way faster than a DateTime could handle. This poses an interesting question as to how to store the books in a database. Right now, when you initialize a new OrderBook a stopwatch is started and that is whats used to determine how many "ticks" have passed and that is what's used in the Order to determine the order in which trades occurred.

Benchmarks

• Dev laptop: i7-3632QM @ 2.2GHz & 16GB:
• Run over 10 seconds.
• Orders submitted: 1,505,542 or 150,554 orders per second
• Trades completed: 1,010,481 or 101,048 trades per second

In Summary

There's a lot that goes into building a high performance exchange but it seems to be doable in .NET as long as the trade volume isn't crazy (like NASDAQ).

TL;DR

Markets have many different kinds of order types, limit being the simplest SortedSet is a .NET implementation of a Red-Black Tree Use it for order books DateTime has terrible precision Use stopwatch instead