Performance evaluation of the OpenACC API targeting multi-GPU data transfers: A case study of porting parallel 3D sweeping algorithms for geological folding

Abstract

The increasing prevalence of GPUs in modern computing presents both performance opportunities and programming challenges. The complexity of heterogeneous CPU+GPU programming and the time-consuming process of adapting existing code to specific architectures often hinder performance optimization. We present a performance evaluation of the OpenACC API for various data transfer methodologies in the context of using multiple GPUs within a single compute node. Through a set of microbenchmarks, we first quantitatively study various scenarios of OpenACC-enabled data communication, encompassing high-level and low-level APIs, with synchronous, asynchronous and P2P variants for the latter. Then, we demonstrate how OpenACC can be used to offload the 3D Parallel Marching Method (3DPMM), a geological folding algorithm employed for mapping the Earth’s crust. When parallelizing the involved 3D stencil computations among multiple GPUs, we need to enable both the standard halo communication across the inter-GPU boundary surface and the non-standard volumetric data rearrangement between the GPUs. A proper use of the OpenACC API is therefore extremely important. To investigate the parallel capabilities of the 3DPMM algorithm, we employ a 3D computational domain and assess grid sizes in each direction ranging from 128 to 1280. Our findings reveal that for grid sizes up to 76 8 3 , a single GPU configuration outperforms setups using 2 and 4 GPUs, which exhibit performance slowdowns of up to 50% and 35%, respectively. In contrast, for problem sizes greater than 76 8 3 , the use of 4 GPUs shows the performance upper hand over 1 and 2 GPUs, achieving an average speedup of approximately 25%.

The authors' abstract, as published at the source. Parallel Computing, 2026 · DOI ↗