What is the hardware that can be used to solve a model given its approximate size and the Feko solver used?

Mel
Mel Altair Community Member
edited July 15 in Altair HyperWorks

Solver scalability refers to how efficiently a solver performs when the number of cores are increased. Solver scalability depends mainly on three things:

  • the hardware specifics (CPU type, cores per CPU, CPUs per node, cluster interconnects)
  • model size (usually referring to the number of mesh elements)
  • solver type (MoM, MLFMM, FEM, …)

In the table below, available hardware can be grouped into three main categories:

  1.  image Notebook or desktop computer
  2.  image Multi-CPU server
  3.  image Compute cluster with multiple computes nodes.

The table lists each of the Feko solvers and categorizes model sizes as "small", "medium" and "large" sizes. Click the MoM and MLFMM for more details.

 

Solver

 

Small

Model Size

Medium

 

Large

MoM

Checkbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fill

 

Checkbox Checked with solid fill

MLFMM

Checkbox Checked with solid fill

 

Checkbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fill

FEM

 

Checkbox Checked with solid fill

Checkbox Checked with solid fill

Checkbox Checked with solid fill

PO

Checkbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fill

RL-GO

Checkbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fill

UTD

Checkbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fillCheckbox Checked with solid fill

Checkbox Checked with solid fillCheckbox Checked with solid fill

FDTD

Checkbox Checked with solid fill

Checkbox Checked with solid fill

Checkbox Checked with solid fill

Regarding solvers and model sizes:

Method of Moments (MoM): This solver scales very well with the number of processes. The solver’s memory requirement will dictate the type of hardware to use. A small model for this solver would have less than 30 000 unknowns (7 GByte of memory) **

Multilevel Fast Multipole Method (MLFMM):  Memory requirements increase with the number of processes. Solver scaling is better when the iterative solver converges in a small number of iterations. A small model for this solver would have less than 300 000 unknowns.

Finite Element Method (FEM): Memory requirements increase with the number of processes. Solver scaling is better when the iterative solver converges in a small number of iterations, but in general does not scale as well as the MoM. A small model for solver would have less than 1 million dofs.

Physical Optics (PO): Solver scales very well, memory requirements mostly dependent on the number of mesh elements*

Ray launching Geometrical Optics (RL-GO): Solver scales very well, memory requirements mostly dependent on the number of mesh elements*

Uniform Theory of Diffraction (UTD): Solver scaling is excellent. Memory requirements do not grow substantially with the number of processes*. Memory requirements depend on the number of UTD surfaces and requested field points.

Finite Difference Time Domain (FDTD): Memory requirements increase with the number of processes. A small model for this solver would have less than 1 million voxels.

* Usually the asymptotic solvers (PO, UTD, RL-GO) are used in combination with the MoM. For example a reflector antenna system could have a horn antenna modelled with MoM and the reflector with PO. The solver scaling will then also depend on the size of the MoM part.

** Unknowns (or degrees of freedom or "dofs" for FEM) are printed near the beginning of the *.out file. It can be obtained by running Feko for a few seconds until after the geometry checking is complete.

Tagged:

Categories