High-level multi-platform approaches for scoring phylogenies on CPU and GPU devices

Abstract

Research on parallel computing has promoted significant advances in the solution of different real-world problems. One of the main application domains that have benefited from the exploitation of parallelism is bioinformatics. However, bioinformatics tools are often constrained by the use of low-level platform-specific programming models. This issue limits the spectrum of hardware that can be used to accelerate time-consuming biological workloads, while also affecting programming productivity and code maintenance. To address such limitations, this work investigates high-level multi-platform approaches to parallelize, as a case study, an important task in evolutionary bioinformatics: the evaluation of phylogenetic quality. Particularly, we define parallel designs of the parsimony scoring function for CPU and GPU devices, using three high-level application programming interfaces with multi-platform support: OpenMP, OpenACC, and SYCL. Different optimizations and algorithmic strategies are defined to boost execution according to the characteristics of the tackled parallel tasks. An in-depth experimental evaluation on real-world data reveals the main strengths and areas of improvement for each high-level multi-platform approach, in comparison to lower-level platform-specific alternatives. In addition, the study of single-source performance portability suggests that CPU designs behave better than GPU ones, yet device-oriented optimizations are still needed to pursue a precise exploitation of computational resources.