Various numerical models have been used to understand and predict ocean dynamics. For this reason, many information technology (IT) resources are required for high-resolution global ocean modeling. The development of cloud-computing technologies has enabled earth scientists to easily use numerical ocean models that require high-performance computing (HPC) and message-passing interface (MPI) software in private and public clouds. Although it is easier today to use computing resources than it was in the past, computational reproducibility and portability in diverse IT environments remain crucial issues. This study proposes a model execution architecture for computational reproducibility, portability, and agility based on container-based virtualization and orchestration technologies. We implement a containerized regional ocean-modeling system (ROMS), an MPI-based numerical ocean model that exists in various public or private cloud environments (e.g., personal computers and multiple-node servers). The preparation time for model setup is greatly reduced using our container-based HPC architecture. Containerization of ROMS is tested for its support of the portability of numerical modeling in a wide range of public-cloud environments. When leveraging an abstraction layer of complex and diverse infrastructure environments, we can run the ocean model more easily while obtaining computational reproducibility using a shareable deployment code. This advancement can be used to guide the containerization of various numerical models and to run them in parallel in public and private cloud-computing environments.

已经使用各种数值模型来理解和预测海洋动力学。因此，高分辨率全球海洋建模需要许多信息技术（IT）资源。云计算技术的发展使地球科学家能够轻松地使用需要在私有云和公共云中使用高性能计算（HPC）和消息传递接口（MPI）软件的数字海洋模型。尽管今天使用计算资源比过去更容易，但是在各种IT环境中计算的可复制性和可移植性仍然是关键问题。这项研究基于基于容器的虚拟化和编排技术，提出了一种用于计算重现性，可移植性和敏捷性的模型执行体系结构。我们实施了集装箱化的区域海洋建模系统（ROMS），在各种公共或私有云环境（例如，个人计算机和多节点服务器）中存在的基于MPI的数字海洋模型。使用我们基于容器的HPC体系结构，可以大大减少模型设置的准备时间。ROMS的容器化已经过测试，可以支持广泛的公共云环境中数值模型的可移植性。当利用复杂多样的基础架构环境的抽象层时，我们可以更轻松地运行海洋模型，同时使用可共享的部署代码获得计算的可重复性。这一进步可用于指导各种数值模型的容器化，并在公共和私有云计算环境中并行运行它们。