Running high resolution coastal models in forecast systems: moving from workstations and hpc cluster to cloud resources

Abstract

Computational forecast systems (CFS) are essential modelling tools for coastal management by providing water dynamics predictions. Nowadays CFS are processed in dedicated workstations, fulfilling quality control through automatic comparison with field data. Recently, CFS has been successfully ported to High Performance Computing (HPC) resources, maintained by highly-specialized staff in these complex environments. The need to increase the available resources for more demanding applications and to enhance the portability for use in non-scientific institutions has promoted the search for more flexible and user-friendly approaches. The scalability and flexibility of cloud resources, with dedicated services for facilitating their use, makes them an attractive option.

Herein, the performance of CFS using ECO-SELFE MPI-based model is assessed and compared for the first time in multiple environments, including local workstations, an HPC cluster and a pilot cloud. The analysis is conducted in a range of resources from the physical core count available at the smaller resources to the optimal number of processes, using cloud and HPC cluster resources. Results for the smaller, common physical resources show that the cloud is an attractive option for CFS operation. As the optimal number of processes for the use case is at the limit of the workstations common pool, an analysis was also performed using HPC cluster nodes and federated MPI resources. Results show that the cloud remains an attractive option for CFS. This conclusion is valid both for the use of a single host or through federated hosts, providing that efficient communication infrastructure (such as SRIOV) is available.