Reduce Your Simulations Time and Costs
Engineers that rely on CFD for advanced design and virtual prototyping strive to increase the accuracy, size, and sophistication of their simulations. They need access to high-performance solvers and high performance computing equipment in order to turn around simulations in a reasonable time-frame.
In the past, accessing this computing capability for CFD applications required purchase or lease of expensive multi-processor systems, along with additional HPC licenses over and above base CFD license fees. Today, new super-computing architectures are making the hardware much more affordable, but other vendors have been slow to respond with a truly economical software package. We built our platform to address these issues and to create a next generation experience.
Our purpose-built manycore CFD solver, coupled with an inexpensive compact CPU/GPU system can outperform traditional commercial solutions available at the same price by an order of magnitude. GPU devices love to be loaded up with tasks. EXN/Aero’s unique parallel task allocation system enables a single CPU/GPU machine to do challenging simulations that would usually require a large multi-CPU system. In this way, parallel computing for CFD ceases to be a costly add-on, and becomes fundamental to software design. EXN/Aero demonstrates that this ideal is possible today.
Everything you need to speed up your engineering simulations.
Steady State & Unsteady State
Synthetic turbulence generation for LES
Unsteady-RANS turbulence modeling
Direct quadrature dispersed phase
Ideal gas models, buoyancy
High speed & compressible flow
Buoyancy driven flows
Hybrid mesh topology – structured / unstructured
Mixed mesh precision – single/double
CGNS database format for input & output
3rd party meshing software compatibility through CGNS format (Pointwise, ICEM Hexa/Tetra, enGrid)
Dynamic load balancing and resource usage optimization
Data output in VTK & CGNS format
A Unique CFD Solution
EXN/View is our interface. It integrates simulation set up & monitoring tools and is designed to work with the CGNS database structure. EXN/View is built on open-source tools and we make the source code freely available to clients & customers that want to pursue their own in-house development or customization.
The image below is the simulation set up utility displaying a mesh for a cylinder in cross-flow.
Our patented cell & interface design subdivides simulation tasks into two classes of objects– cells and interfaces. This allows us to be much more flexible and creative in the parallelization scheme.
- Cell objects contain the CFD mesh and represent the bulk of the computational load. Cells might contain structured or unstructured mesh and use single or double precision.
- Interface objects are multifunction data tranformation utilities with many uses, including boundary condition input, cell-to-cell communication, reduction of output data, addition of source terms and interpretation of real-time sensor inputs.
In a manycore computing environment, all devices operate on tasks concurrently, and not all architectures are equally good at every computing task.
Task allocation considers the strengths & weaknesses of different processor types. Bulky cell operations are often computed on GPU devices where the thousands of simple processors and fast memory make short work of the job. CPU devices coordinate the simulation and handle the more complex interface tasks.
Resources are not bound to a single task and so multiple simulations can run simultaneously on a single CPU/GPU resource. This concept improves productivity for teams of engineers sharing a compute resource, and lies at the heart of our space & time parallelization.
Most CFD solutions enable users to reduce processing time by using multiple processors. The domain is divided in as many chunks as there are available processing cores, and each core is given a chunk to compute. Spatial decomposition speeds things up compared to serial processing, but is expensive to scale up.
Our space-time parallelism approach divides the domain in both space and time. Several instances of a spatially-parallelized domain are started simultaneously on the same CPU/GPU machine, where each instance is a “time slab” of the total simulation time window. Time slabs are computed concurrently, and communicate with each other via a predictor-corrector algorithm. An extra multiplier of 2x – 3x is possible using this layered parallelism approach.
As always, we believe users shouldn’t have to pay to work on challenging CFD problems. Space-time parallelization is a standard feature in EXN/Aero.