Use this link to cite:
https://hdl.handle.net/2183/49343 CUDA Parallelization of the SGBM Method for Swing Option Pricing
Loading...
Identifiers
Publication date
Authors
Adwani, Bhavna
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Abstract
[Resumo] Este traballo presenta o desenvolvemento e a avaliación dun acelerador GPU para o método de agrupación de redes estocásticas (SGBM) aplicado á valoración de opción swing en mercados enerxéticos. A implementación secuencial en C é validada fronte aos valores de referencia publicados, perfilada con gprof e perf para identificar os colos de botella computacionais, e a continuación portada a CUDA cunha arquitectura de cinco kernels que paraleliza a simulación Monte Carlo, a clasificación de traxectorias e a regresión polinomial local. Os experimentos realízanse na GPU NVIDIA A100 do supercomputador Finis Terrae III do CESGA, e os resultados mostran unha aceleración máxima de 20.3× respecto á liña base secuencial con 𝑁 = 220 traxectorias, cunha precisión financeira do 0.03% respecto ao valor de referencia publicado para 𝑁 = 4 × 106 traxectorias e 𝜈 = 1024 bundles.
[Abstract] This thesis presents the development and benchmarking of a GPU-accelerated implementation of the Stochastic Grid Bundling Method (SGBM) for pricing swing options in energy markets. A sequential C implementation is validated against published reference values, profiled using gprof and perf to identify computational bottlenecks, and subsequently ported to CUDA using a five-kernel architecture that parallelises the Monte Carlo simulation, path bundling via thrust::sort_by_key, and local polynomial regression. Experiments are conducted on the NVIDIA A100 GPU of the CESGA Finis Terrae III supercomputer. Results demonstrate a peak speedup of 20.3× over the sequential baseline at 𝑁 = 220 paths, with financial accuracy within 0.03% of the published reference value at 𝑁 = 4 × 106 paths and 𝜈 = 1024 bundles.
[Abstract] This thesis presents the development and benchmarking of a GPU-accelerated implementation of the Stochastic Grid Bundling Method (SGBM) for pricing swing options in energy markets. A sequential C implementation is validated against published reference values, profiled using gprof and perf to identify computational bottlenecks, and subsequently ported to CUDA using a five-kernel architecture that parallelises the Monte Carlo simulation, path bundling via thrust::sort_by_key, and local polynomial regression. Experiments are conducted on the NVIDIA A100 GPU of the CESGA Finis Terrae III supercomputer. Results demonstrate a peak speedup of 20.3× over the sequential baseline at 𝑁 = 220 paths, with financial accuracy within 0.03% of the published reference value at 𝑁 = 4 × 106 paths and 𝜈 = 1024 bundles.
Description
Editor version
Rights
Attribution-NonCommercial-NoDerivatives 4.0 International







