A Scalable Dataflow Implementation of Curran's Approximation Algorithm
Anna Maria Nestorov, Enrico Reggiani, Hristina Palikareva, Pavel Burovskiy, Tobias Becker, Marco Domenico Santambrogio · 2017
Computational finance is a challenging application domain with ever-increasing performance requirements. Driven by the competition between companies, computational finance pushes High Performance Computing (HPC) technology to its limits. In this paper, we consider Asian options which are financial derivatives whose payoff is determined by the average price of their underlying asset at predetermined observation points rather than on the single value at expiration time. Due to this path dependency, their pricing is computationally expensive and is therefore a suitable candidate for dataflow acceleration. This paper introduces an application for Asian option pricing based on Curran's approximation method that exploits a dataflow-oriented development approach, employing dedicated optimisations and replacing conventional floating-point with fixed-point formats wherever possible. The implementation targets a Maxeler server-class HPC system consisting of a CPU server node and Maxeler dataflow engines encapsulating Altera Stratix V FPGAs. The application has been evaluated on two different data sets and achieves a speed-up of 111x and 278.3x compared to a single-threaded software implementation, and 4x and 9.2x compared to a multi-threaded software implementation running on a dual socket CPU server with 12-core Intel Xeon E5-2697 v2 CPUs with up to 48 hyper-threads in total.