混合异构众核平台上的高性能计算金融算法研究

发布时间：2018-06-03 16:24

本文选题：高性能计算 + 期权定价　；参考：《山东大学》2017年硕士论文

【摘要】：从计算机诞生开始,人们就在不断的追求更高的计算速度,高性能计算一直是计算机研究领域重要的研究内容。目前高性能计算的主流思路是采用多核并行的方式,并且随着异构计算的兴起,利用异构加速设备提高计算能力成为高性能计算的主流方法,越来越多的超级计算机采用了异构加速的结构来获得更高的计算能力。高性能计算也在金融,生物,地理勘探,图像处理,气象等众多应用环境下发挥出了越来越重要的作用。相较于传统计算,异构计算能够显著提高整体的计算能力,使得复杂算法能够的到实现和应用,使得很多复杂算法得到应用。在大数据时代的现在,异构计算提供了处理超大规模数据的解决方法。在金融领域里,算法实时性要求很高,很多复杂的算法由于计算时间过长限制了其应用场景。应用高性能计算能够很好的解决复杂算法的时效性问题,为算法应用提供了一种很好的解决方法。期权作为一种重要的金融衍生物,其定价算法具有非常好的代表性。蒙特卡洛算法能够通过对价格进行模拟从而给出期权价格,是一种重要的期权定价的数值算法,也因为其不需要显示表达式的特点,也被应用于高维期权定价等问题,取得很好的效果,但蒙特卡洛算法需要模拟非常多的路径才能得到较为精确的算法,其计算时间过长导致算法应用受到限制。对于时序数据,RNN算法能够很好的发掘时间中的规律,其中采用LSTM模型的RNN结构对于长时间记忆具有较好的效果,目前越来越多的被应用于语言模型,自然语言处理,金融时间序列分析,视频分析等具有明显时间特征的数据分析和预测中。本文在详细介绍期权背景和定价算法的基础上,分别针对欧式期权和美式期权,使用蒙特卡洛算法进行计算,并针对不同的蒙特卡洛算法特征提出对应的异构并行算法,并在混合异构平台上加以实现,能够实现同时利用多种异构众核进行加速,并取得了很好的加速效果。本文详细对比了不同的异构设备在加速时的不同特征,并针对不同设备给出了优化方法。为了能够更好的利用节点上全部的计算能力,多种平台下的异构计算框架,能够同时支持CPU,MIC,GPU混合架构,并能实现跨节点的计算,获得了很好的节点扩展性。对于金融市场价格走势的预测对于分析金融时间序列和指导投资具有重要意义,本文选择外汇数据,引入LSTM模型的RNN神经网络算法,对LSTM模型的分析,实现了一个能够对价格走势进行预测的时序模型,使用了 Adam训练方法,对外汇价格的涨跌进行了预测。为了能够更好的应用高性能计算,对LSTM模型的算法进行改进,通过分析LSTM算法模型的热点情况,将热点计算函数做了优化。在本文提出了新的针对LSTM算法的高性能数学加速库MPL(Math Parallel Library)。使用MPL数学库,能够迅速完成LSTM模型在多种异构设备下的并行加速。对MPL数学库使用OpenCL进行加速,并针对Intel Intergrated Graphics Card进行优化,达到了很好的加速效果。
[Abstract]:Since the birth of the computer, people are constantly pursuing higher computing speed. High performance computing has always been an important research content in the field of computer research. At present, the mainstream idea of high performance computing is a multi core parallel way, and with the rise of heterogeneous computing, it is high performance to use heterogeneous acceleromers to improve computing power. More and more supercomputers use heterogeneous accelerated structures to obtain higher computing power. High performance computing also plays a more and more important role in many applications, such as finance, biology, geographic exploration, image processing, meteorology and so on. Compared with traditional computing, heterogeneous computing can significantly improve the overall performance. The computational ability makes the complex algorithm can be implemented and applied, and many complex algorithms are applied. In the large data age, the heterogeneous computing provides a solution to the large scale data processing. In the financial field, the real-time performance of the algorithm is very high, and many complex algorithms have limited the application due to the long time of calculation. The application of high performance computing can be a good solution to the timeliness of complex algorithms, and provides a good solution for the application of the algorithm. As an important financial derivative, the option pricing algorithm has a very good representation. An important numerical algorithm for option pricing is also applied to the problem of high option pricing because it does not need to display the characteristics of the expression. But the Monte Carlo algorithm needs to simulate a very large number of paths to get a more accurate algorithm, and its calculation time is too long to limit the application of the algorithm. Time series data, RNN algorithm can well discover the rules of time, in which the RNN structure of LSTM model has a good effect on long time memory. At present, more and more are applied to language model, Natural Language Processing, financial time series analysis, video analysis and other data analysis and prediction with obvious time characteristics. On the basis of detailed introduction of option background and pricing algorithm, the Monte Carlo algorithm is used to calculate the European option and American option, and the corresponding heterogeneous parallel algorithms are proposed for the different Monte Carlo algorithm features. It can be realized on the mixed heterogeneous platform and can be implemented simultaneously with a variety of heterogeneous crowd cores. In order to better utilize all the computing power on the node, the heterogeneous computing framework under various platforms can support the CPU, MIC, GPU hybrid architecture at the same time. It is of great significance to predict the financial market price trend and to guide investment in the analysis of the financial market price trend. This paper selects foreign exchange data, introduces the RNN neural network algorithm of the LSTM model, analyzes the LSTM model, and realizes the prediction of the price trend. The time series model uses the Adam training method to predict the rise and fall of the foreign exchange price. In order to better apply the high performance calculation, the algorithm of the LSTM model is improved. The hot computing function of the LSTM algorithm is optimized by analyzing the hot situation of the LSTM algorithm model. A new high performance mathematics for the LSTM algorithm is put forward in this paper. The acceleration library MPL (Math Parallel Library). Using the MPL math library, the parallel acceleration of the LSTM model can be speeded up in a variety of heterogeneous devices. The MPL mathematics library is accelerated with OpenCL, and the Intel Intergrated Graphics Card is optimized to achieve a good acceleration effect.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP38

【相似文献】