Optimization strategies for portable code for Monte Carlo-based value-at-risk systems

Abstract

Value-at-risk (VaR) computations are one important basic element of risk analysis and management applications. On the one hand, risk management systems need to be flexible and maintainable, but on the other hand they require a very high computational power. In general, accelerators provide high speedups, but come with a limited flexibility. In this work, we investigate two approaches towards portable and fast code for VaR computations on heterogeneous platforms: operator tuning and the use of OpenCL. We show that operator tuning can save up one third of run time on CPU-based systems in the calibration step. For OpenCL, we present a detailed analysis of run time on CPU, GPU, and Xeon Phi, and evaluate its portability. We also find that the same code runs up to 12x faster in a VaR setting with an accelerator card being present, without any code changes required.