Performance evaluation of unified memory and dynamic parallelism for selected parallel CUDA applications

The aim of this paper is to evaluate performance of new CUDA mechanisms--unified memory and dynamic parallelism for real parallel applications compared to standard CUDA API versions. In order to gain insight into performance of these mechanisms, we decided to implement three applications with control and data flow typical of SPMD, geometric SPMD and divide-and-conquer schemes, which were then used for tests and experiments. Specifically, tested applications include verification of Goldbach's conjecture, 2D heat transfer simulation and adaptive numerical integration. We experimented with various ways of how dynamic parallelism can be deployed into an existing implementation and be optimized further. Subsequently, we compared the best dynamic parallelism and unified memory versions to respective standard API counterparts. It was shown that usage of dynamic parallelism resulted in improvement in performance for heat simulation, better than static but worse than an iterative version for numerical integration and finally worse results for Golbach's conjecture verification. In most cases, unified memory results in decrease in performance. On the other hand, both mechanisms can contribute to simpler and more readable codes. For dynamic parallelism, it applies to algorithms in which it can be naturally applied. Unified memory generally makes it easier for a programmer to enter the CUDA programming paradigm as it resembles the traditional memory allocation/usage pattern.