| Media Contact: Tim Niu, Marketing Manager Phone: +1-781-273-3322 Email: tim@comsol.com Reader Contact: |
Media Contact, Europe: Bernt Nilsson, VP of Marketing Phone: +46-8-412 95 17 Email: bernt@comsol.com |
Download this press release in Word format
Download high-resolution images
Benchmarks of Scientific Modeling Software
Reveal That FEMLAB Rivals Specialized Packages
Two university-based research groups completed benchmark testing of three leading packages-FEMLAB, ANSYS and Fluent. The results show that flexibility and ease of use need not amount to a sacrifice in performance.
Stockholm, Sweden (April 30, 2004) - Scientific-modeling software packages conventionally are optimized to quickly and efficiently solve certain suites of problems within a particular field of specialty. One might assume (errantly) that speed and accuracy would be lost when one flexible and easy-to-use software package is employed to tackle single physics problems. Thanks to a recent focus on high-performance computation from the COMSOL development team, a new multiphysics release, FEMLAB 3.0a, achieves the same speed and accuracy numbers of niche scientific packages. Independent benchmark tests confirm that FEMLAB 3.0a closes the speed and accuracy gap that previously separated multiphysics and niche software classes.
A general tool as nimble as a specialized package
FEMLAB is a multiphysics package capable of modeling any physical phenomena that can be described with partial differential equations (PDEs). It is this emphasis on solving PDEs that makes FEMLAB effective within and between widely ranging disciplines. "We saw no reason that being general and flexible should mean giving up the efficiency and computational performance that is attained by software packages designed specifically for solving one-type physics problems," explains Svante Littmarck, CEO of COMSOL Inc. "After all, the heart of modeling a problem in structural analysis, fluid flow, electromagnetics, or acoustics is solving a PDE."
FEMLAB uses state-of-the-art solvers developed primarily within the academic research community to address both single- and multi-physics problems. Most dedicated physics packages employ similar solvers. "Because these codes share the same core algorithms for the computationally intensive tasks, " continues Littmarck, "we believed we could make our general-purpose package just as fast, accurate, and memory efficient as specialized programs." Littmarck adds, "We used results of previous benchmark tests to help identify the bottlenecks in our code that we resolved in our most recent version, FEMLAB 3.0a."
FEMLAB uses state-of-the-art solvers developed primarily within the academic research community to address both single- and multi-physics problems. Most dedicated physics packages employ similar solvers. "Because these codes share the same core algorithms for the computationally intensive tasks, " continues Littmarck, "we believed we could make our general-purpose package just as fast, accurate, and memory efficient as specialized programs." Littmarck adds, "We used results of previous benchmark tests to help identify the bottlenecks in our code that we resolved in our most recent version, FEMLAB 3.0a."
Benchmark testing and reports
Benchmark testing involves creating the same model in different software packages and comparing the memory use, solution times, and accuracy obtained with each model. Ideally the benchmarking is conducted by independent analysts using scientific methods in a non-biased framework. The FEMLAB 3.0a benchmark evaluation was conducted by two independent research groups-The Parallel and Scientific Computing Institute at The Royal Institute of Technology (Stockholm, Sweden) and the Centre for Mathematical Sciences at the Lund Institute of Technology (Lund, Sweden). The two groups compared FEMLAB 3.0a with dedicated-physics packages, ANSYS and Fluent.
The testing centered on well-known problems in which clear and definite parameters exist as reference metrics. Structural mechanics capabilities of FEMLAB and ANSYS were evaluated using standardized problems from NAFEMS (National Agency for Finite Element Methods and Standards) as well as examples directly from the ANSYS manuals. Fluid-dynamics capabilities of FEMLAB and Fluent were tested using classic models from the scientific literature.
Results of the testing have been reported and are available to the public. In the reports, examiners define problems, explain procedures, and provide data with sufficient detail for readers to reproduce and verify the results. The complete academic reports are available on the Internet:
The testing centered on well-known problems in which clear and definite parameters exist as reference metrics. Structural mechanics capabilities of FEMLAB and ANSYS were evaluated using standardized problems from NAFEMS (National Agency for Finite Element Methods and Standards) as well as examples directly from the ANSYS manuals. Fluid-dynamics capabilities of FEMLAB and Fluent were tested using classic models from the scientific literature.
Results of the testing have been reported and are available to the public. In the reports, examiners define problems, explain procedures, and provide data with sufficient detail for readers to reproduce and verify the results. The complete academic reports are available on the Internet:
- Benchmark of FEMLAB, Fluent and ANSYS
by Olivier Verdier, Lund Institute of Technology, Centre for Mathematical Sciences and Numerical Analysis http://www.maths.lth.se/na/staff/olivier/BenchmarkReport2.pdf - Benchmarking FEMLAB 3.0a: Laminar Flows in 2D
by Michael Hanke, Royal Institute of Technology (Stockholm), Department of Numerical Analysis and Computer Science, Parallel and Scientific Computing Institute. http://www.psci.kth.se/Activities/Reports/Results/2004/benchm.pdf
Bridging the gap
The following tables give an overview of selected results from the benchmark studies. Parameters of particular interest are accuracy, execution time, and peak memory usage. The studies document the speed and memory consumption needed to reach a prescribed accuracy for each test problem. They present the accuracy figure as the negative logarithm of the relative deviation from a reference value - the higher value, the better the accuracy. An accuracy value 1.0 corresponds to 90% of the exact value; a value of 2.0 equals 99%, 3.0 equals 99.9%, and so on. The number of DOF (degrees of freedom) measures the problem size.
For this model, the size and CPU time are roughly equal between FEMLAB and the single-physics product, but FEMLAB consumes less memory to achieve higher accuracy.
Here problem sizes are almost identical. Compared with its competitor, FEMLAB achieves the same accuracy for principal stress, greater accuracy for displacement, and far less memory consumption with roughly comparable CPU times.
Problem sizes in this comparison are roughly equal. While the FEMLAB model consumes greater memory than the single-physics package, the FEMLAB results are far more accurate and require less computing time than the single-physics product.
"We are very proud to bring our general-purpose multiphysics modeling software to performance levels scientists and engineers expect from specialized programs," says Littmarck. "Benchmark testing is a great value for the engineering community and COMSOL in particular. We look forward to continuing with this type of testing in our development of FEMLAB in the near future."
Structural Mechanics:
Elliptic membrane; linear elastic analysis of stress in the y-direction.| Program | Number of DOF(thousands) | Peak Memory (MB) | CPU time(seconds) | Y-direction Stress Accuracy |
|---|---|---|---|---|
| Ansys 7.1 | 74 | 180 | 10 | 2.67 |
| FEMLAB 3.0a | 76 | 135 | 9 | 3.12 |
For this model, the size and CPU time are roughly equal between FEMLAB and the single-physics product, but FEMLAB consumes less memory to achieve higher accuracy.
Structural Mechanics:
Built-in plate; linear elastic analysis of displacement and principal stress.| Program | Number of DOF(thousands) | Peak Memory (MB) | CPU time (seconds) | Displacement Accuracy | Principal Stress Accuracy |
|---|---|---|---|---|---|
| Ansys 7.1 | 101 | 547 | 72 | 1.22 | 1.05 |
| FEMLAB 3.0a | 101 | 309 | 85 | 1.38 | 1.07 |
Here problem sizes are almost identical. Compared with its competitor, FEMLAB achieves the same accuracy for principal stress, greater accuracy for displacement, and far less memory consumption with roughly comparable CPU times.
Fluid Dynamics:
Laminar flow around a cylinder in 2D. The cylinder is slightly offset from the center of the channel. This results in a lifting force, measured as the lift coefficient. The drag force of the cylinder is measured as the drag coefficient.| Program | Number of DOF(thousands) | Peak Memory (MB) | CPU time (seconds) | Drag Coefficient Accuracy | Lift Coefficient Accuracy |
|---|---|---|---|---|---|
| Fluent 6.1.18 | 109 | 67 | 450 | 1.97 | < 1 |
| FEMLAB 3.0a | 101 | 371 | 108 | 4.75 | 2.13 |
Problem sizes in this comparison are roughly equal. While the FEMLAB model consumes greater memory than the single-physics package, the FEMLAB results are far more accurate and require less computing time than the single-physics product.
"We are very proud to bring our general-purpose multiphysics modeling software to performance levels scientists and engineers expect from specialized programs," says Littmarck. "Benchmark testing is a great value for the engineering community and COMSOL in particular. We look forward to continuing with this type of testing in our development of FEMLAB in the near future."
About FEMLAB
FEMLAB - which stands for Finite Element Modeling Laboratory - is an advanced software package for the modeling and simulation of any physical process described with PDEs. The latest version, FEMLAB 3.0a, features high-performance, state-of-the-art solvers that address extremely large problems yet quickly yield accurate results. Working in an easy-to-use graphical interface or from a command line, users choose from several ways to describe their problems in 1D, 2D, and 3D. A particular strength of the package is its PDE modeling capability, enabling equations from various fields such as structural mechanics, electromagnetics, fluid flow, and chemistry to be linked and solved all in the same model and all at the same time. These and many other features make FEMLAB 3.0a an unprecedented modeling environment for research, product development and education.
About COMSOL
Founded in 1986 in Stockholm, Sweden, COMSOL has grown to include U.S. offices in Burlington, MA, and Los Angeles, CA. Internationally the firm has operations in Denmark, Finland, Norway, Germany, France and the United Kingdom. Full details about the company and its products are available at www.comsol.com.
