Open Access
Subscription Access
Heterogeneous Parallelism of High-Order Residual Distribution Schemes Using Central and Graphics Processing Units
Abstract
The parallelism of a high-order residual-distribution algorithm was investigated on a heterogeneous system consisting of a multi-core CPU and a graphics processing unit (GPU). Solutions of the steady-state Euler equations were obtained by executing a logistically simple, but computationally expensive, portion of the algorithm on the GPU while the logistically complex, but relatively inexpensive, remainder is simultaneously computed on the CPU cores. The resulting hybrid parallelism, combined with the massive efficiency of the GPU, conceals the computationally expensive portion and the overall speedup is shown to be defined by the fraction executing on the CPU cores, according to Amdahl’s law. Based on the observed speedup, the monetary savings provided by a heterogeneous CPU and GPU architecture over that of only CPU cores are estimated.