Navigation and service

Fire Simulations on High Performance Computing

The application of numerical simulation software on high performance computing (HPC) systems comes with great benefits and challenges. Ideally, the execution time of the model reduces linearly with the amount of used resources. This assumes, that the software is developed for parallel processing. In the case of FDS, it was developed as a serial application and parallelism was added afterwards. Therefore the FDS model can be executed on distributed memory machines. The following figure illustrates the parallel performace of FDS for each submodel and the total execution time (main routine).

FDS ScalingFDS MPI Scaling – the total execution time is represented by the main curve

While MPI was added at the beginning of the last decade, the OpenMP parallelisation was added more recently. The current OpenMP implementation is based on the work done by Haarhoff and Arnold (2014). It allows the execution of FDS on shared memory systems and in a hybrid mode, i.e. a combination of MPI and OpenMP. As the current OpenMP implementation allows for only a speed up of about 2 on 4 or more cores, see following figure.

FDS ScalingFDS OpenMP Scaling - scalability for three problem sizes is shown

The usage of FDS on HPC systems requires additional scripts, e.g. automated simulation restarts and batch system scripts. We provide support for all users of JSC's facillities as well as others on request. One of the developed tools is FDSgeogen that supports automated scenario and geometry generation for FDS.

The simulations in this group are mostly carried out on the JURECA supercomputer under the computing time grants jjsc11 and jjsc27.

Publications

Performance Analysis and Shared Memory Parallelisation of FDS
D. Haarhoff and L. Arnold
Proceedings and Talk. Fire and Evacuation Modelling Technical Conference 2014

Theses

Hybrid Parallelisation of the CFD-software FDS
Daniel Haarhoff
Master thesis, 2014