high performance computing on graphics processing units: hgpu.org

We present a novel parallel binomial algorithm to compute prices of American options. The algorithm partitions a binomial tree into blocks of multiple levels of nodes, and assigns each such block to multiple processors. Each processor in parallel with the others computes the option’s values at nodes assigned to it. The computation consists of two phases, where the second phase can not start until the valuation in the first phase has been completed. The algorithm is implemented and tested on a heterogeneous system consisting of an Intel multicore processor and a NVIDIA GPU. The whole task is split and divided over the CPU and GPU so that the computations are performed on the two processors simultaneously. In the hybrid processing, the GPU is always assigned the last part of a block, and makes use of a couple of buffers in the on-chip shared memory to reduce the number of accesses to the off-chip device memory. The performance of the hybrid processing is compared with an optimised CPU serial code, a CPU parallel implementation and a GPU standalone program. We learned from the experiments that the lack of explicit mechanism in CUDA for synchronising CPU and GPU executions is a major obstacle for the hybrid processing to achieve high performance.