Field Precision adds support for parallel processing to Aether

Field Precision has added support for parallel processing to Aether, a 3D code for time and frequency-domain electromagnetic simulations. The measured speed improvement is close to a factor of 2X on a dual-core computer.

Parallel processing has long been a promising path to significant speed increases in computer simulations. Recent technological developments make parallel processing a practical reality on personal computers running Windows:

• The ability to address effectively unlimited memory in Windows XP64 and Vista64.
• The growing prevalence of dual-core and quad-core machines.

Field Precision has the goal of supporting automatic parallel processing in the 64-bit versions of all of our software packages by mid-2010. Because the programs are written in FORTRAN 95 with careful optimization, the serial versions already achieve impressive speed. The clean code organization makes parallelization straightforward and effective. With parallel processing, the company expects speed amplification close to 2X for a dual-core machine and 4X for a quad-core machine.

Field Precision recently created a parallel version of Aether, a 3D time and frequency-domain electromagnetic code. The unified finite-element package handles all stages of a solution: mesh generation, field calculation and interactive analysis. Time-domain calculations are appropriate for pulsed-power devices, electromagnetic interference from transient events and the diffusion of magnetic fields into conductive structures. In the frequency domain, Aether handles both closed and open structures. Closed system tasks include searches for resonant frequencies, generation of mode field distributions and calculation of $Q$ factors. Open-system solutions include antenna simulation and the determination of S parameters for microwave networks.

Benchmark tests were performed with unmodified examples from the Aether application library. In a resonant-mode search, the standard serial version of the program completed the task in 255 seconds. On the average, the task used one processor of a dual-Xeon system. The parallel version used 100% of the system capability and completed the task in 142 seconds. The run time was 56% of the serial-program time, close to the ideal.