An explicit dynamics GPU structural solver for thin shell finite elements

Question

1. What are the contributions mentioned in the paper "An explicit dynamics gpu structural solver for thin shell finite elements" ?

2. What future works have the authors mentioned in the paper "An explicit dynamics gpu structural solver for thin shell finite elements" ?

3. What is the main objective of this work?

4. What is the only step that requires explicit communication between different elements?

Accepted Answer

Cremonesi et al. this paper presented an optimized, double-precision finite element GPU implementation of an explicit dynamics finite element solver for elastic shell problems in small strains and large displacements and rotations, using unstructured meshes.

Accepted Answer

This work constitutes a starting point for the future solution of more complex problems.. Another important topic of future developments will be the introduction of contact algorithms in the GPU implementation and the possibility of simulating fracture propagation.. With all these ingredients, the code will be able to guarantee effective computing times ( near real-time in some special cases ) for the simulation of engineering problems of real use.

Accepted Answer

The main objective of this work is to develop a highly efficient explicit finite element structural solver for the simulation of the dynamical behavior of thin-walled structures.

Accepted Answer

In the standard time stepping algorithm with central difference method, the force calculation step is the only one that needs explicit communications: after the nodal forces are computed for each element separately, they need to be exchanged between neighboring elements, in order to compute the right nodal force contributions.

Accepted Answer

The proposed dynamic load balancing algorithm allows for the efficient use of both CPUs and GPUs, minimizing the idle time and exploiting the capabilities of the heterogeneous hardware.

Accepted Answer

Another advantage of the proposed element-wise algorithmic structure is that when an element is considered, the only mesh connectivity information necessary to the CUDA kernel for each one of the element local nodes concerns the indexes of neighboring elements which share it and the corresponding local node index inside those elements.

Accepted Answer

The nonlinear structural problem is discretized by means of MITC4 shell elements [20, 21], where the typical constraints of Reissner-Mindlin shell kinematics are enforced.

Accepted Answer

From earlier tests with CUDA profiling tools, it emerged that in the presented CUDA kernel the biggest cause of performance drop is caused by register spilling: there were too many local variables to fit the available registers, thus some of them were being stored in local memory (which has the same latency as global memory) causing a significant drop in performance.

Accepted Answer

In addition, the solver currently runs on single-GPU machines, even though extending the current implementation to a multi-GPU environment is straightforward.

Accepted Answer

In fact, runningthe simulations with single-precision arithmetic, speedups of 46x and 61x are obtained considering Abaqus adaptive and fixed time step configurations respectively.

Accepted Answer

In this way, a single CUDA kernel is sufficient to perform all calculations concerning an entire time step, avoiding multiple kernel calls, with the relative overhead, and minimizing memory accesses.

Accepted Answer

A recent review of applications in FEM (Finite Element Method)-based structural mechanics can be found in [1], but GPUs are being used also in a variety of other contexts.

Accepted Answer

a certain performance drop when using double-precision instead of single-precision in GPUs has to be accepted as unavoidable (even though on newer GPUs this gap is being reduced, on older ones double-precision could only be emulated and thus performances were very poor).

Accepted Answer

To make results easier to compare, the GPU thread-block size auto-tuning feature is turned off: all blocks are fixed to 32×32 threads.

An explicit dynamics GPU structural solver for thin shell finite elements

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What are the contributions mentioned in the paper "An explicit dynamics gpu structural solver for thin shell finite elements" ?

2. What future works have the authors mentioned in the paper "An explicit dynamics gpu structural solver for thin shell finite elements" ?

3. What is the main objective of this work?

4. What is the only step that requires explicit communication between different elements?

5. What is the proposed dynamic load balancing algorithm?

6. What is the advantage of the proposed algorithmic structure?

7. How is the nonlinear structural problem discretized?

8. What is the main reason for the performance drop in the CUDA kernel?

9. What is the current implementation of the solver?

10. How many times are the speedups obtained with Abaqus adaptive?

11. What is the way to perform the calculations on a single CUDA kernel?

12. What is the recent review of applications in FEM?

13. What is the reason why the performance drop is not a problem?

14. How is the GPU thread-block size auto-tuning?

Figures

Citations

Efficient topology optimization using GPU computing with multilevel granularity

A GPU-based DEM-FEM computational framework for tire-sand interaction simulations

A Class-Kriging Predictor for Functional Compositions with Application to Particle-Size Curves in Heterogeneous Aquifers

Simulating direct shear tests with the Bullet physics library: A validation study

The interactions between an off-road tire and granular terrain: GPU-based DEM-FEM simulation and experimental validation

References

Theory of plates and shells

Finite Element Procedures

Nonlinear Finite Elements for Continua and Structures

Finite Element Analysis

Theory of Plates and Shells

Related Papers (5)

GAMER: a GPU-Accelerated Adaptive Mesh Refinement Code for Astrophysics

Development of parallel explicit finite element sheet forming simulation system based on GPU architecture

Accelerating CFD simulation with high order finite difference method on curvilinear coordinates for modern GPU clusters.

An Efficient GPU-Based Out-of-Core LU Solver of Parallel Higher-Order Method of Moments for Solving Airborne Array Problems

Particle transport with unstructured grid on GPU