Investigating and reducing latency of trading applications
Institutions that implement high-frequency trading may gain a significant competitive advantage in the open market by outperforming market trends. This is achieved with careful design and implementation of trading algorithms who's latency can be managed in a deterministic way.
However, there still remain external sources of latency:
- multiple application threads contending for the same CPU core
- suboptimal scheduling with respect to NUMA layout
- unmanaged interrupt processing
- unnecessary OS features left enabled that require processing time
- and others
This presentation will cover methods that were used to discover external latency based on both conventional sources of information available in Linux and emerging technologies: perf, ftrace and systemtap.
Devops engineers, system administrators, application developers of low-latency or real-time applications who are interested in minimizing and managing latency. Also everyone interested in linux kernel and new kernel tracing technologies.
I work as a devops engineer at Deutsche Bank Tech Centre. We create technology for trading on financial markets. My main focus is providing stable Linux platform, solving performance issues is a part of the work. Previously I worked at MCST, producer of CPU Elbrus which is based on VLIW architecture, where I ported linux distribution and investigated performance issues of software on new platform.