Products: Abaqus/Standard Abaqus/AMS
Benefits: New scalable thread-parallel execution capability of the AMS eigensolver significantly improves the performance of frequency extraction analyses.
Description: Scalable thread-parallel execution of the AMS eigensolver delivers significant performance improvement on shared memory computers and on a single node of a computer cluster.
Table 6–1 illustrates the improved performance of the AMS eigensolver on a system with Intel Nehalem processors for two industrial models: Model 1 is a 4.3 million degree-of-freedom automotive powertrain model with a large selective recovery node set and damping projection, and Model 2 is a 9.2 million degree-of-freedom automotive vehicle model with a large selective recovery node set. The wall-clock times in the table indicate the total elapsed times for the frequency extraction step using the AMS eigensolver.
Table 6–1 AMS performance improvement due to a new scalable thread-parallel execution capability.
| Model | Degrees of Freedom (Millions) | Abaqus 6.10 (1-core) | Abaqus 6.10-EF (4-core) | ||
|---|---|---|---|---|---|
| Number of Modes | Wall Clock Time (h:mm) | Number of Modes | Wall Clock Time (h:mm) | ||
| Model 1 | 4.3 | 1709 | 1:02 | 1710 | 0:29 |
| Model 2 | 9.2 | 4208 | 2:00 | 4210 | 1:03 |
Large-scale models specifying full recovery of eigenmodes may require a considerable amount of physical memory to avoid extra I/O operations, which may lead to a degradation of parallel scaling. Due to this enhancement and the approximate nature of the AMS technology, it is possible to observe slight differences in the number of eigenmodes extracted by AMS in Abaqus 6.10-EF versus Abaqus 6.10. These differences are expected since AMS eigenmodes close to the user-specified maximum frequency are generally less accurate and more sensitive to perturbations (e.g., changes in the order of the system of equations). However, the results of subsequent modal dynamic procedures are very close to the results in Abaqus 6.10 and previous releases if an appropriate number of modes are used to construct the projection basis.
Parallel scaling of the AMS eigensolver is improved in Abaqus 6.11 for multi-core shared memory computers having more than 4 cores. Table 6–2 illustrates the improved parallel scaling of the AMS eigensolver for a 9.3 million degree-of-freedom automotive vehicle model with selective recovery. The AMS eigensolver computes 7326 eigenmodes for this model.
Table 6–2 AMS performance improvements due to improved parallel scaling on multi-core systems.
| Number of Cores | Abaqus 6.10-EF | Abaqus 6.11 | ||
|---|---|---|---|---|
| Wall Clock Time (sec) | Speedup | Wall Clock Time (sec) | Speedup | |
| 1 | 5650 | 1.00 | 5367 | 1.00 |
| 4 | 2560 | 2.21 | 2257 | 2.38 |
| 8 | 2185 | 2.25 | 1762 | 3.05 |
| 16 | 2149 | 2.63 | 1604 | 3.35 |
In addition, the residual mode computation is parallelized in Abaqus 6.11 to improve the performance for models with many residual vector nodes. The performance of the overall full recovery procedure is improved for the case where the full recovery requires a large number of I/O operations.
References: