Corvil » White Papers
White Papers
Analysis for Low Latency Trading Networks
The world’s financial markets are abandoning the telephone and the trading floor as their means of transacting business. Trading now takes place predominantly in interconnected computer networks where orders are placed and matched via machines. Increasingly, orders also originate from machines running sophisticated automatic trading algorithms. In this new environment, the lifespan on which data retains relevance has shrunk to the machine’s timescale of milliseconds or less. Simultaneously, the volume of market data distributed by trading systems has grown steadily as automated algorithms make heavier use of quotes.
These trends have challenged the providers of trading systems to achieve ever-lower latency and handle ever-larger traffic volumes. A key requirement for achieving predictable high performance is the ability to measure and track the performance of the infrastructure in place. One potential performance bottleneck is the network linking the electronic traders to the pricing and execution servers in the data centres. Traditional network measurements, which take place over timescales of seconds and minutes, provide little or no insight into behaviour at milli- and microsecond timescales. But these are the timescales which now determine the performance boundary of a trading network.
Accurate measurement of network latencies in these ranges, in a live environment, presents its own set of unique challenges. Issues such as timestamp latency and clock synchronization can sabotage the unwary. And care must be taken to ensure that measured latency and loss correspond to what is actually experienced by application traffic. Activity on the network occurs in bursts, which cause queues to build up and buffers to overflow, and generally create their own latency conditions. Visibility into these short-lived conditions is central to understanding network performance. Measurements of both service levels and bit rates must be able to resolve events that appear and disappear in milliseconds. And once the measurements have been collected, analyzing them will also require new tools and techniques, which can reveal the short-timescale links between latency, traffic behaviour, and network resources.
