MultiThreading
SOFA has some multithreading capabilities in its core, but more features are available in the MultiThreading plugin.
Plugin Compilation
- Enable
PLUGIN_MULTITHREADINGin your CMake configuration. It is disabled by default. - Configure, Generate and build.
Parallel Collision Detection
Most SOFA scenes use a component defining the collision pipeline. This pipeline requires two components for the broad phase collision detection, and the narrow phase collision detection. A usual choice for the broad phase is BruteForceBroadPhase, and BVHNarrowPhase for the narrow phase.
In the MultiThreading plugin, two components can replace BruteForceBroadPhase and BVHNarrowPhase.
ParallelBruteForceBroadPhase
This component is a parallel implementation of BruteForceBroadPhase using a global thread pool. It means the result of a simulation with BruteForceBroadPhase or with ParallelBruteForceBroadPhase is expected to be equal. ParallelBruteForceBroadPhase is the most efficient compared to BruteForceBroadPhase when there is a lot of objects in the scene.
Examples
Examples of ParallelBruteForceBroadPhase can be found in:
ParallelBVHNarrowPhase
This component is a parallel implementation of BVHNarrowPhase using a global thread pool. It means the result of a simulation with BVHNarrowPhase or with ParallelBVHNarrowPhase is expected to be equal.
Examples
Examples of ParallelBruteForceBroadPhase can be found in:
Physics
Parallel FEM
ParallelTetrahedronFEMForceField
ParallelTetrahedronFEMForceField is the multi-threaded equivalent of TetrahedronFEMForceField.
This implementation is the most efficient when the number of tetrahedron is large (> 1000).
The following methods are executed in parallel: – addDForce – addKToMatrix
Examples
Examples of ParallelTetrahedronFEMForceField can be found in:
ParallelHexahedronFEMForceField
ParallelHexahedronFEMForceField is the multi-threaded equivalent of HexahedronFEMForceField.
This implementation is the most efficient when:
- the number of hexahedron is large (> 1000)
- the global system matrix is not assembled. It is usually the case with a CGLinearSolver templated with GraphScattered types.
- the method is ‘large’. If the method is ‘polar’ or ‘small’,
addForceis executed sequentially, butaddDForcein parallel.
The following methods are executed in parallel:
addForcefor method ‘large’.addDForce
The method addKToMatrix is not executed in parallel. This method is called with an assembled system, usually with a direct solver or a CGLinearSolver templated with types different from GraphScattered. In this case, the most time-consuming step is to invert the matrix. This is where efforts should be put to accelerate the simulation.
Examples
Examples of ParallelHexahedronFEMForceField can be found in:
BeamLinearMapping_mt
This component inherits all the functionality from the BeamLinearMapping component and overrides three virtual functions that contain a for loop: apply(), applyJ() and applyJT(). It adds only one data attribute, the granularity. This attribute sets the number of iterations of the for loop, corresponding to the number of points along the beam elements that must be assigned and executed for each task. If this number is lower than the number of iterations the loop won’t be parallelized, and the corresponding BeamLinearMapping function is called. If this number is greater than the number of iterations of the loop the tasks are created, and each task executes the granularity value of iterations of the loop.
Examples
Examples of BeamLinearMapping_mt can be found in:
Springs
ParallelStiffSpringForceField
ParallelStiffSpringForceField is the multi-threaded equivalent of StiffSpringForceField.
ParallelMeshSpringForceField
ParallelMeshSpringForceField is the multi-threaded equivalent of MeshSpringForceField.
Examples
Examples of ParallelMeshSpringForceField can be found in:
Independent SOFA Scenes
The AnimationLoopParallelScheduler component was implemented to run the physics simulation of independent scenes in parallel. The component looks for BaseAnimationLoop components in all its child nodes and executes the step() function of each AnimationLoop in parallel.
The DataExchange component can manage the sharing of data between all the concurrent scenes without being bound by synchronization locks. To avoid the use of synchronization locks, each component in different scenes must have its own copy of the same data to share, and the data synchronization is executed serially. After the data synchronization the VisualLoop is executed in serial. Each child node of the node where the AnimationLoopParallelScheduler is placed must be seen as an independent scene and there should be no physics interaction or collision detection between these scenes. When all the step() functions return, the visual loop (graphics rendering) is executed serially throughout the scene.
The DataExchange component must be placed in the same node where the AnimationLoopParallelScheduler is, and the path links to the source and destination data to copy must be defined. The data template of the source and destination path links must be the same. This data synchronization is executed serially before the VisualLoop is executed. A common use of the AnimationLoopParallelScheduler component is to place it in the root node of the scene hierarchy and add a AnimationLoop to all the child nodes you want to be executed independently.
Tips and Limitations
To get the best performance with the AnimationLoopParallelScheduler component, the execution time of each animation loop should be manually balanced to get almost the same time length. If in the scene an animation loop is computationally more expensive, the MultiStepAnimationLoop component can be useful not to limit the speed of execution of a potentially faster AnimationLoop step function. This will improve the synchronization between threads, minimizing the waiting time for all animation loops completion. The scene should be split into as many independent scenes as possible. An independent scene is considered as a scene where the physics simulated objects are never supposed to collide with the objects of another scene during the whole simulation time length.
The main limitation using the AnimationLoopParallelScheduler is that interaction with the mouse with the objects in the scene crashes the simulation.
Examples
Examples of AnimationLoopParallelScheduler and DataExchange can be found in:
Last modified: 3 March 2023
