Stop Conditions#

Depending on the model under investigation, the required number of iteration steps may vary widely. Let’s say we are in a scenario where we are investigating the convergence behaviour of a model; depending on the parameter configuration, the time to reach convergence could span many orders of magnitude.

In order to save computational resources, it is useful to dynamically stop simulations, depending on their current state.

The stop conditions feature implements exactly that: depending on a timeout or on information conveyed to the frontend using the monitor method, running simulations can be told to stop.

Using stop conditions#

Configuring a simulation run#

To configure such a stop condition, add an entry like the following to your run configuration:

  # The total timeout (in seconds) for finishing _all_ simulations
  timeout: !expr 60*60*24   # 24h in this case. Default: None

  # A list of StopCondition objects to check during the run _for each worker_.
  # The entries of the following list are OR-connected, i.e. it suffices that
  # one is fulfilled for the corresponding worker to be stopped.
    # First condition (using the shortest possible syntax)
    - func: timeout_wall
      seconds: 0.4
      # further arguments to utopya.stopcond_funcs.timeout_wall

    # Second condition (for showcasing the long syntax)
    - name: single simulation timeout (in long syntax)
      description: terminates a worker if it ran for too long
      enabled: true
        # The functions specified here are AND-connected, i.e.: _all_
        # need to return True for the stop condition to be fulfilled
        - func: timeout_wall
          # will look for utopya.stopcond_funcs.timeout_wall method
          seconds: 0.4
          # further arguments to utopya.stopcond_funcs.timeout_wall

    # Third condition: Terminating depending on a monitor entry
    - name: extinction
      description: terminates a simulation if entity density goes to zero
        - func: check_monitor_entry
          entry_name: MyModel.density   # TODO set your monitor entry name here
          operator: <=
          value: 0.
          # Will stop if: {monitor entry} <= 0.

This example showcases three registered stop conditions. The first two implement the same condition, a timeout. They do so once in short and once in long syntax (the short syntax is useful if to_check contains only a single function).

The third condition checks against the current model state, as conveyed to the frontend via its monitor method.

Currently, the following stop condition functions are available in the stop_conditions module:

  • timeout_wall() terminates a single universe if it took too much time (measured against the clock on the wall, not CPU time)

  • check_monitor_entry() reads the latest monitor entries and compares it to some value using an operator.


If you are missing a stop condition function, please let us know by opening an issue or a merge request via the project page.


To check if a monitor entry is in an interval, pass multiple entries to the to_check argument; these are AND-connected. Example:

    # Will stop if: 41.9 <= {monitor entry} <= 42.1
    - !stop-condition
      name: fixation
      description: terminates once reaching the known fixpoint at 42
        - &check_density
          func: check_monitor_entry
          entry_name: MyModel.density
          operator: '>='
          value: 41.9
        - <<: *check_density
          operator: '<='
          value: 42.1


Of course, stopping a simulation can also be implemented directly on C++ side. However, with the stop conditions feature you will have more flexibility and do not need to implement additional logic into your model; you only need to provide the monitor entries and set the configuration entries.

During the simulation run#

Let’s say that we choose the following run configuration:

perform_sweep: true

  # The global timeout
  timeout: 100  # in seconds

  # The stop conditions: stop after 10s or if the mean state gets too large
  # Can specify this with or without YAML tag
    - func: timeout_wall
      seconds: 10
    - !stop-condition
      name: large CA state
      description: stops simulation when the CA's maximum value passes 13
      func: check_monitor_entry
      entry_name: ExtendedModel.ca_max
      operator: ">="
      value: 13

  forward_streams: true

  # Use many steps to ensure that the simulation runs long enough; also, do not
  # want to write data that frequently ...
  num_steps: 10M
  write_every: 100k

  # Use a small monitor emit interval such that the information is quickly
  # available.
  monitor_emit_interval: 0.3

    model: info

  seed: !sweep
    default: 42
    range: [13]

    state_size: 10000

The log output will then show that it detected both the timeout and the stop conditions:

PROGRESS multiverse      Adding tasks for simulation of 13 universes ...
INFO     multiverse      Added 13 tasks.
PROGRESS workermanager   Preparing to work ...
NOTE     workermanager     Timeout:         13:37:42 (in 1m 40s)
NOTE     workermanager     Stop conditions: timeout_wall, high state
HILIGHT  workermanager   Starting to work ...

During the rest of the simulation, everything is as always. If the stop conditions kick in, the run should be finished more quickly, though.


There will be no extra log output if a universe was stopped. Also, the progress bar will show the status of the run in the same way as without stop conditions: it does not distinguish between universes that finished due to a stop condition from those that finished regularly after num_steps.


If the stop conditions are not detected by the WorkerManager, make sure that the run_kwargs end up in the meta configuration. You can do so by inspecting the config directory in the output folder, where all configuration parts are backed up separately.

After the simulation run#

To find out which universes were stopped due to a stop condition, inspect the _report.txt file in the output directory. It will contain information on which stop conditions were fulfilled for which universes.

In the example shown above (using the dummy model and a purely illustrative set of stop conditions) all simulations end up being stopped due to the high state stop condition rather than the wall timeout.

Stop Conditions

  13 / 13 universes were stopped due to at least one of the following stop conditions:

  StopCondition 'timeout_wall'

  StopCondition 'high state': stops simulation when state gets too high
      uni01, uni02, uni03, uni04, uni05, uni06, uni07, uni08, uni09, uni10, uni11, uni12, uni13


The _report.txt file is already created during the run and updated when individual simulations end. If you are interested in that information before the end of the simulation run, you can also inspect it during the run.

Conveying information from your model to the frontend#

As mentioned above, the check_monitor_entry function can make use of the values of a monitor entry and use those to stop a simulation.

For an implementation example, let’s look at the Forest Fire model’s monitor method:

void monitor () {
    this->_monitor.set_entry("tree_density", calculate_tree_density());

Further information on how to set monitor entries can be found in the doxygen documentation.


The monitor method is only invoked after certain time intervals, independent of the number of iteration steps. As these are invoked quite seldom, it usually is no big issue if you perform more costly operations within monitor.

To control how frequently the monitor data is emitted and thus becomes available to the frontend, set the parameter_space.monitor_emit_inveral entry.

We suggest not choosing values that are significantly smaller than one second. This is because all monitor data is retained by the frontend, thus potentially eating up your memory if the monitor is emitting too frequently or the simulation time is very long.


What exactly happens once a stop condition is fulfilled?

If the frontend detects that a stop condition for a specific task (i.e., a universe) is fulfilled, the following procedure takes place:

  1. The frontend sends the SIGUSR1 signal to the running simulation

  2. The Model base class catches that signal and sets a flag, denoting that the simulation should stop.

  3. The model will finish its current iteration step (i.e., the iterate call).

  4. The model will then quit with a non-zero exit code.

  5. The corresponding code is caught by the frontend, allowing to detect that the simulation ended due to a stop condition. Unlike for regular non-zero exit codes, no verbose log messages are printed, nor are exceptions raised.

  6. Some tracking variables are updated to reflect this information.

  7. At the end of the simulation, the _report.txt shows information on which universe was stopped due to which stop condition.