The Application Of High Performance Computing In The Risk Computing Area Of Interest-rate Products

During the continuous reform and innovation of China's financial market,the volatility of the value of financial products is unpredictable,the market is becoming more and more open,and the risks are also growing.With the increasing number of financial market participants and more and more rich financial products,the computational resources required for financial risk calculation will also gradually increase.Therefore,we attempt to apply high performance computing technology,especially GPU heterogeneous computing technology,to the field of financial risk calculation of interest rate products,which enables it to utilize high performance computing resources to improve its computing performance and also improve the efficiency and the accuracy of the financial risk assessment.In this paper,we mainly concerned with the calculation of the VaR and the fitting of the term structure of the interest rate products.The four commonly used algorithms are: BDT-based pricing model,historical VaR algorithm,cubic spline interpolation algorithm and smoothing B-spline algorithm.Algorithm parallelization,GPU porting and performance optimization for algorithm features and GPU architecture are carried out.For computationally intensive and abundant financial risk computing tasks,we designed task-level parallelism and finer-grained parallelism for all four algorithms to fully utilize the computing power of multi-core GPU architecture.In addition,combining the algorithmic features with the hardware features of the GPU,our optimization work on the ported program includes:(1)The use of Shared Memory.Such as putting the frequently accessed data into Shared Memory and using Shared Memory as a cache.(2)Coalesced memory access.By adjusting the memory layout and thread task allocation,the number of memory access operations can be reduced by coalesced memory access.(3)Multi-stream optimization.When the number of tasks is large,tasks can be divided into pieces and thus GPU computing tasks can be started in multiple batches,so as to cover part of the data preparation and transmission time.After porting and optimizing work,all four algorithms have achieved higher performance on the NVIDIA K80 GPU than the dual E5-2693v3 CPU platform.Compared with the single-threaded CPU program,all four algorithms achieve more than 41 x speedup;Compared to the multi-threaded CPU program,all four algorithms achieve a speedup of 2.1x or more.Our experiment results show that the financial risk computing of interest rate products can achieve higher performance through GPU-based heterogeneous acceleration.