Main Content

Run Co-Simulation Components on Multiple Cores

Simulink® supports co-simulation between components using local solvers or that involves simulation tools. For example, co-simulation can involve an S-function implemented as a co-simulation gateway between Simulink and third-party tools or custom code. It can also involve an FMU in co-simulation mode imported to Simulink or a Model block in accelerator mode.

To improve performance, consider running models that contain co-simulation blocks (components) on multiple threads if:

  • You are integrating multiple co-simulation components

  • Integration at the component level is computationally intense

This topic assumes that you are familiar with multithreaded programming and concepts.

Simulink lets you run C MEX S-functions, Co-Simulation FMU blocks, and Model blocks configured for accelerator mode on multiple threads if they are threadsafe. Being threadsafe means that the block can work with multiple threads accessing shared data, resources, and objects without any conflicts.

You can also run a model on multiple threads if the model has a nondirect feedthrough block followed by a direct feedthrough block (see Singlethreaded Versus Multithreaded Simulations). Direct feedthrough means that the output port signal of a block is computed from the values of its input port signals in the same time step.

Note

Multithreaded co-simulation for Level-2 MATLAB S-Function blocks is not supported.

By default, Simulink configures all models to run on multiple threads. If performance can be improved, Simulink automatically runs all models on multiple threads. However, not all models have co-simulation components that can run on multiple threads, and not all models benefit from running on multiple threads.

Singlethreaded Versus Multithreaded Simulations

When evaluating a model for possible concurrencies in your system for multithread execution, Simulink uses a graph-based algorithm to analyze data dependencies among eligible blocks and form multiple task groups. This algorithm ensures that the different task groups:

  • Have no data dependencies on each other

  • Can execute in parallel using multiple cores

For example, Simulink might look for non-direct feedthrough blocks followed by direct feedthrough blocks. If the following direct feedthrough blocks depend only on a nondirect feedthrough block, the software can group the blocks in the same branch for parallel execution.

This example compares the serial execution of a series of blocks partitioned for parallel execution.

graph-based execution multithread conceptual diagram

Using Multithreaded Simulation to Speed up Simulation on Multiple Cores

Consider the Graph-Based Multithread Simulation example, which uses four nondirect feedthrough blocks followed by four direct feedthrough blocks. Direct feedthrough means that the output port signal of a block is computed from the values of its input port signals in the same time step.

Simulink analyzes the data dependencies and forms four branches. Each branch contains nondirect feedthrough blocks followed by direct feedthrough blocks. Then, Simulink uses multithreading to run those four branches. In this way, Simulink can speed up the simulation up to four times on a four or more core machine compared to a singlethreaded simulation on the same machine. You can manually measure the simulation of the model before and after changing the MultithreadedSim parameter.

slexGraphBasedMultiThreadSimExample model

You can manually measure the simulation of the model before and after changing the MultithreadedSim parameter. This code shows how to measure simulation for the Graph-Based Multithread Simulation example:

openExample('slexGraphBasedMultiThreadSimExample');...
set_param(gcs,'MultithreadedSim','off');
tic;sim('slexGraphBasedMultiThreadSimExample');toc
Elapsed time is 39.406743 seconds.
set_param(gcs,'MultithreadedSim','on');
tic;sim('slexGraphBasedMultiThreadSimExample'); toc
Elapsed time is 13.619744 seconds.

In this example, the speedup is almost three times (instead of four times) on a four core machine. This is due to the extra overhead introduced by the operating system thread context switch.

Using the MultithreadedSim Parameter

You can specify that an entire model run on multiple threads, or specify that particular blocks run on multiple threads, using the MultithreadedSim parameter. Specify that an entire model runs on multiple threads if all the co-simulation blocks in the model are threadsafe. If some, but not all blocks, are threadsafe, identify only those blocks to run on multiple threads. The model and blocks use the MultithreadedSim parameter as follows:

SettingDescription

'auto'[a]

(Default) Let Simulink decide if the block can run on multiple threads.

'off'

Disable the block or model from running on multiple threads.

[a] Starting in R2021a, when you open a model created in a previous release, Simulink interprets the 'on' setting to be the same as 'auto'.

Enabling the MultithreadedSim parameter does not mean that the block or model simulates on multiple threads. Simulation on multiple threads occurs when MultithreadedSim is enabled and:

  • The block and/or model operate at a single rate.

  • The block and/or model are threadsafe. (For example, they do not use static or global data).

  • The block and/or model is exception-free. For S-function blocks, use the ssSetOptions function to set SS_OPTION_EXCEPTION_FREE_CODE.

Multithreading does not allow solver reset checks, and therefore skips over any use of the ssSetSolverNeedsReset and ssBlockStateForSolverChangedAtMajorStep functions. Conversely, in accelerator mode, if these functions are used or there are continuous states, multithreading is automatically turned off.

Configuring S-Function Blocks to Run Single or Multithreaded

Whether an S-function block runs single or multithreaded depends on the MultithreadedSim parameter value and the ssSetRuntimeThreadSafetyCompliance function.

MultithreadedSim SettingssSetRuntimeThreadSafetyCompliance SettingSingle or Multithread
'auto'RUNTIME_THREAD_SAFETY_COMPLIANCE_UNKNOWNSingle thread
'auto'RUNTIME_THREAD_SAFETY_COMPLIANCE_TRUEMultithread
'auto'RUNTIME_THREAD_SAFETY_COMPLIANCE_FALSESingle thread
'off'The setting is ignored and the S-function block runs singlethreaded

Co-Simulation on Multiple Threads Limitations and Guidelines

  • The simulation runs on a single thread for accelerator and rapid accelerator modes. Multithreading is enabled when the simulation mode is normal.

    Set model simulation mode to normal.

  • There is no code generation for co-simulation components.

  • Multithreading is not activated for blocks with constant sample time.

  • Multithreading is not enabled when the Simulink debugger is on.

    Turn off Simulink debugger.

  • A block that depends on a non-thread-safe block cannot be multithreaded. Consider breaking the dependency, for example, by using a Unit Delay block.

S-Function Block Limitations

  • Must have a single rate.

    Consider revising your model to break down multirate components into individual single-rate components.

  • Multithreading is not enabled when an S-function has variable sample time.

    Consider using a different sample time (see Specify Sample Time).

  • Multithreading is not enabled when an S-function has continuous states and solver is fixed-step, which together trigger a continuous states consistency check. To disable continuous states consistency checks, use the ssSetSkipContStatesConsistencyCheck function.

  • Must be thread safe — In ssSetRuntimeThreadSafetyCompliance(SimStruct *S,int_T val), val must be RUNTIME_THREAD_SAFETY_COMPLIANCE_TRUE.

    For more information, see Guidelines for Writing Thread-Safe S-Functions.

  • Must be exception-free — In ssSetOptions(SimStruct *S,uint_T options), options must include SS_OPTION_EXCEPTION_FREE_CODE.

    For more information, see Guidelines for Writing Thread-Safe S-Functions.

  • Multithreading is not enabled when the S-function Analyzer is on. Try multithreading in normal mode.

  • Multithreading is not enabled when S-function has continuous sample time. Consider using a different sample time (see Specify Sample Time).

  • Multithreading concurrently runs output and update methods. The block must have an output or update method.

FMU Import Block Limitations

  • Must be in co-simulation mode.

    Consider switching FMU mode from Model Exchange to Co-Simulation.

  • Must be thread-safe, for example, multiple FMUs must not access the same file at the same time.

  • Multithreading is not enabled when FMU block logging displays in the MATLAB® command window. Redirect FMU block logging to a file using:

    set_param(blockName,'FMUDebugLoggingRedirect','File')
  • Multithreading is not supported when FMU is running out-of-process. To disable this setting, use:

    set_param(blockName,'DebugExecutionForFMUViaOutOfProcess','off')

Model Block Limitations

  • Multithreading is not enabled when a Model block has event ports.

  • Cannot be inside a For Each Subsystem block.

    Consider moving the Model block out of the For Each Subsystem block.

  • Must be in accelerator mode.

  • Must have single rate.

    Consider revising your model to break down multirate components into individual single-rate components.

  • Cannot use blocks with variable sample time.

    Consider using a different sample time (see Specify Sample Time).

  • Cannot have continuous states.

  • Consider breaking the dependency between blocks, for example, by using a Unit Delay block.

  • Must have a fixed-step solver.

  • Cannot access any global data stores.

  • Multithreading is not enabled when a model contains a Simulink Function block.

  • Cannot use any Simulink functions or caller blocks.

  • Cannot contain To File blocks. For more information, see Export Simulation Data.

  • Cannot contain From File blocks.

    Consider feeding data into the referenced model via an inport from the top-level model.

See Also

| | |

Related Topics