Numerically intensive simulations on an integrated computing infrastructure

Numerically intensive simulations on an integrated computing infrastructure

Scientific simulations are usualy very computationally intensive and time consuming and can benefit greatly by a suitable selection and mapping to specialized computing resources. In modern data centers with heterogeneous computing infrastructures, such as cloud services, high performance computing (HPC) clusters and specialized computing resources with accelerators (for example GPU) this choice is not trivial (see Figure 1). Therefore in this project concepts and technical implementations of a user-transparent integration mechanism in a heterogeneous computing infrastructure should be developed. This should unify and simplify the access to specialized computing resources for computationally intensive simulation projects within the SWZ. Then a simulation application is to be semi-automated executed on the most appropriate resource type.

As an application of these concepts the simulation of transport and transformation processes in porous materials is used. For this calculations  highly efficient Lattice Boltzmann methods are very suitable. This application forms the basis for detailed performance analysis and an appropriate modeling.

Goals of the project

In the project scientists of the research fields Fluid Dynamics (research group Prof. Brenner, TU Clausthal), Software Engineering (research group Prof. Grabowski, University of Göttingen) and the Applied Computer Science (research group of Prof. Yahyapour, GWDG) work together to develop appropriate solutions. Additional support in the development of numerical methods for the application will be received from the research group of Prof. Deiterding from Southampton University.

From an engineering point of view the aim of this project is to develop an understanding of material transport and transformation in porous materials. The background of the project is the increasing demand in various fields of engineering, for example, chemical engineering, to design porous materials and to use them  for example as supports for catalysts with improved conversion rates and selectivity. Therefore it is necessary to be able to predict the morphology-transport relationships involving chemical reactions more accurately than previously possible.

For the mathematically point of view the main task is the development of suitable numerical methods. The starting point for this are works which are already performed by the cooperation partner in the field of computational fluid dynamics, in particular the lattice Boltzmann (LB) method, the adaptive mesh refinement (see Figure 2) and the parallel and high performance computing and cloud computing. Until now there are only some raw theoretical models for calculation of multiphase fluids in such structures which are based on the LB method. To use this method for calculation of technically relevant porous structures, further developments are needed to dissolve the processes running on very different length scales in numerical methods.

From informatics point of view, the interest is to facilitate access to heterogeneous computing infrastructures for Simulation Sciences and to integrate them together. The starting point are the heterogeneous data center infrastructure resources of the Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen (GWDG), which can be considered as a prototype for many data centers at university level. In addition to an HPC cluster and an IaaS cloud GWDG also offers computing resources with specialized accelerator cards. The aim is to develop and implement concepts and solutions for an integrated computing environment for simulation, which decreases the decision for a specific resource type simulation scientists and can execute a simulation application as optimally as possible.

The concept of an integrated computing environment is shown in Figure 3. On application layer, the application is defined by a model, which is then passed to a distribution controller, mapping the application to a particular type of resource (the resource layer), for example, a cloud system, a HPC cluster, a GPU cluster or a grid system. Here, the application model is transformed into a resource-specific input model for the target system. This entry model is then passed to the resource manager of the target system that is running the application on the corresponding infrastructure.

In the context of SWZ project 11.4.1 currently the SimPaaS component is being developed to automatically deploy and run simulation applications in a cloud environment. It is intended to integrate SimPaaS into the framework, which is to be developed in this project.