Performance Optimisation of Atmospheric Model ECHAM6 through Component Concurency

Abstract

This dissertation aims at providing a solution and support for optimizing the performance of the atmospheric model ECHAM6. The special prominence of this research is due to the application of the model in the German climate modeling initiative PalMod for simulating a complete glacial cycle from the last interglacial to the Anthropocene. The model, however, suffers from poor scalability at low resolution, as used in this paleoclimate study, due to the limited number of grid points. As a consequence, the potential of the existing high-performance computing architectures cannot be utilized for such experiments at full scale. Endeavors to adopt a higher optimized model is, thus, opportune for the PalMod research. Our investigation reveals that radiative transfer is a relatively expensive atmospheric process in ECHAM6, accounting for approximately 50% of the total simulation time. This current level of cost is achieved by performing radiation calculations only once every two simulation hours. In response, this dissertation reports on a twofold research effort to alleviate such a computational burden in order to render the paleoclimate simulations viable. It first presents the idea of the concurrent radiation scheme for extending the available concurrency in ECHAM6 further by running the radiation component in parallel with other atmospheric processes. This solution also offers a way forward to improve the accuracy of the simulations by increasing the physical consistency between atmospheric states. To implement this scheme, a novel program analysis approach, i.e. Component Isolation, is then introduced for performing the following tasks: extracting a component from a Fortran program, detecting the input and output global variables of the component, making shared source code of the component non-shared. This approach not only benefits the implementation of the concurrent radiation scheme, but it also provides support for another optimization solution envisaged in PalMod. Furthermore, the high accuracy of the approach of Component Isolation and the implementation of the concurrent radiation scheme is demonstrated using careful qualification and validation. Moreover, a thorough analysis is investigated to show the impact of the new radiation scheme on the performance of the atmospheric model ECHAM6. The experiments show that ECHAM6 can achieve a speedup of over 1.9× using the concurrent radiation scheme while becoming almost twice more scalable. The scientific results from the new scheme are, however, evaluated through an independent investigation by a climate scientist. In the nutshell, the approach of Component Isolation offers an unprecedented solution for preparing any arbitrary components of scientific models (in Fortran) for a better organization scheme such as component concurrency or any individual optimization technique such as mixed-precision arithmetic in order to improve the performance of the model. In addition, the successful example of the concurrent radiation scheme in the atmospheric model ECHAM6 can encourage similar optimization research in scientific computing whenever scalability becomes challenging.