Numerical simulation have emerged in the last decades as a very valuable tool to design better products thanks to the evolution of the numerical methods on one hand, but mostly thanks to the evolution of computing power. At Free Field Technologies, we are working in close collaboration with the leading actors of the High Performance Computing landscape to make sure that Actran is always at the forefront of technical evolution in this domain.
Diverse HPC options to fit each Actran calculation and each infrastructure
Actran features a large number of solutions to improve the computation performances. This grants Actran users a great flexibility so each calculation can run in the most efficient way on their particular hardware. These solutions include:
- Shared and distributed memory parallel processing (frequency, loadcase, axisymmetry order or domain parallelism)
- Optimized MPI and BLAS libraries
Large calculations may require a large amount of RAM to run. When the amount of RAM on a single computing node is not enough, domain parallelism can be used to spread the calculation on several machines with lower amount of RAM.
On the contrary, when the amount of RAM on a machine is just sufficient, multi-threading can be used to boost the performances of a calculation by using threads that are not used by any parallel process. The computation time is reduced while the memory consumption remains the same. Multi-threading has been significantly improved in recent releases of Actran.
In order to maximize Actran's performance we offer dedicated HPC consulting services and we guide you through the Actran setup and optimization on your high-performance computing infrastructure.
GPU computing for Actran DGM
Nowadays, graphic processing units (GPUs) that were used for computer graphics are able to perform computations that usually were addressed by CPUs. Recent computations GPUs (such as NVidia K40, K80 or P100) provide thousands of cores and started accompanying CPUs as co-processors for increasing the computation power of workstations and servers. However the powerfulness of GPUs is plainly achieved for codes with highly intense and local arithmetic.
Actran DGM is a perfect candidate for GPU acceleration thanks to its high order representation of the solution inside elements and its solution discontinuity between elements. A deep code restructuring has been done and the CUDA framework has been integrated. This results in an accelerated version of Actran DGM for NVidia GPUs.
For more information, please read the paper Efficient Aeroacoustic Simulations of Jet Engine Noise using GPU Acceleration.
High performance solvers for all applications
Actran features a large variety of efficient solvers to solve each Actran calculation in the most efficient way:
- Direct solvers (MUMPS, PARDISO)
- Iterative solver (CG_ILU)
- Multifrequential Krylov solver
- Staggered solver
Free Field Technologies is an active member of the MUMPS consortium. This grants us access to the latest upgrades of the MUMPS solver before is released and allows us to suggest improvements for future versions of the solver. Thanks to this early access to new release, we are able to integrate new capabilities in the software as soon as they are released and make sure that we are making the fullest use of the MUMPS solver.
The integration of MUMPS single precision solver is a very good example of these benefits. For most Actran Acoustic and Actran TM cases the use of MUMPS single precision will divide by two the computation time and memory consumption without any effect on the results.
Getting the most out of new processor technologies
Free Field Technologies collaborates closely with Intel to make sure that Actran is using as well as possible the computational resources brought by the latest generations of Intel processors. Each new generation of Intel processor is thoroughly tested and Actran is optimized to run efficiently on each new architecture. The gain in performance brought by each new generation of processor can be up to 25% in the factorization step of the Actran calculation.
Actran also includes the high-performance Intel MKL Pardiso solver which is very memory efficient and therefore very useful to solve large symmetric or unsymmetric matrix problems.