QCon is a practitioner-driven conference designed for technical team leads, architects, and project managers who influence software innovation in their teams.
Graham Markall, TweetApplied Maths Developer at OpenGamma
Biography: Graham Markall
Graham Markall is an Applied Maths Developer at OpenGamma. He works on the OG-Maths library, which uses native code generation to provide an environment for developing fast and robust numerical methods from within Java.
Prior to joining OpenGamma, he completed a PhD at Imperial College London, where he developed a runtime compilation framework for implementing high-performance solvers for partial differential equations. The software tools that he developed as part of his research are being used to re-engineer the Fluidity flow-modelling application, so that it can run on next-generation high-performance architectures.
Twitter: @gmarkall
Presentation: TweetCatastrophic Cancellation: the Pitfalls of Floating-point Arithmetic
The design of IEEE754 floating-point arithmetic makes it safe to use in most cases without specialised knowledge of numerical analysis. However, it only approximates arithmetic on real numbers, leading to hazards that can produce wildly inaccurate results.
This talk covers the common problems that can occur, with a view to avoiding them in practice. I will discuss how IEEE754 approximates real numbers, and present some common patterns in which inaccuracy can occur. I will provide an overview of some manual and automated floating-point optimisations, and how they can safely be used.