High Performance Computational Financial Algorithms On Heterogeneous Many Core Architecture

From the birth of the computer,people are constantly in the pursuit of higher computing speed,High-Performance Computing(HPC)has been an important research field of computer research.In recent years,with the rise of heterogeneous computing,the use of hybrid accelerator to improve the computing power has become a mainstream of high-performance computing method,more and more supercomputers use hybrid accelerators to obtain higher computing power.HPC also plays an increasingly important role in many applications such as finance,biology,geography,image processing,meteorology and so on.Compared with the traditional calculation,heterogeneous computing can significantly improve the overall computing power,and making a lot of complex algorithms applied.In the era of large data now,heterogeneous computing provides a solution to deal with very large data.In the financial field,the real-time evaluation of the algorithm is a very important indicator.Many complex algorithms limit their application scenarios due to their long computation time.The application of high performance computing can solve the problem of the duration of complex algorithms,which provides a good solution for the application of algorithms.Option as an important financial derivative,the pricing algorithm is a good representation.Monte Carlo algorithm can simulate the price so that to calculate the option price,which is an important numerical algorithm.For the time series data,the RNN algorithm can well explore the law of time.The RNN algorithm with LSTM model has good effect on long time memory,and it is often applied to data with obvious time characteristics such as language model and natural language processing processing.Based on the detailed introduction of the option background and the pricing algorithm,this paper uses the Monte Carlo algorithm to calculate the European options and American options respectively,and proposes the corresponding heterogeneous parallel algorithm for the different Monte Carlo algorithm.Implemented on different accelerator platform to harness a variety of heterogeneous accelerators simultaneously,and achieved a very good acceleration effect.In order to make better use of all the computing power on the node,heterogeneous computing framework of a variety of platforms,supporting CPU,MIC,GPU hybrid architecture,and achieved cross-node computing with a good node scalability.For the analysis of financial market price trend for the analysis of financial time series and guide the investment is of great significance.I use foreign exchange data in the LSTM model RNN neural network algorithm,to achieve a price trend can predict the timing Model,using the Adam training method to improve memory capacity,the ups and downs of foreign exchange prices were predicted.In order to improve the performance of the LSTM model,the hotspot calculation function is optimized by analyzing the hotspot of the LSTM algorithm model.The numerical calculation is encapsulated into the MPL math library,and the MPL math Library,can quickly complete the LSTM model in a variety of heterogeneous devices under the parallel acceleration of the MPL math library using OpenCL to accelerate,and optimized for Intel integrated graphics,to achieve a very good acceleration effect.