Define how Dakota should run a function evaluation

Specification

Alias: none

Argument(s): STRINGLIST

Required/Optional	Description of Group	Dakota Keyword	Dakota Keyword Description
Optional		analysis_components	Provide additional identifiers to analysis drivers.
Optional		input_filter	Run a pre-processing script before the analysis drivers
Optional		output_filter	Run a post-processing script after the analysis drivers
Required (Choose One)	Group 1	system	(Not recommended) Launch analysis drivers with a system call
		fork	Launch analysis drivers using fork command
		direct	Run analysis drivers that are linked-to or compiled-with Dakota
		matlab	Run Matlab through a direct interface - requires special Dakota build
		python	Run Python through a direct interface - requires special Dakota build
		scilab	Run Scilab through a direct interface - requires special Dakota build
		grid	Deprecated grid computing interface
Optional		failure_capture	Determine how Dakota responds to analysis driver failure
Optional		deactivate	Deactivate Dakota interface features for simplicity or efficiency

Description

The analysis_drivers keyword provides the names of one or more executable analysis programs or scripts, a.k.a. "drivers" which comprise a function evaluation. The optional and required sub-keywords specify how Dakota will manage directories and files, and run the driver(s).

Types of Interfaces

Dakota has two recommended ways of running analysis drivers:

as an external processes (fork), or
using internal code to couple to the analysis driver (direct)

Other options are available for advanced users, and are not as well documented, supported, or tested:

external processes (system)
internal coupling (python, matlab, scilab, grid)

Use Cases

The internally coupled codes have few options because many of the details are already handled with the coupling. Their behavior is described in the direct keyword.

For external processes using the fork keyword,

A function evaluation may comprise:

A single analysis driver: Function evaluation, including all pre- and post-processing is contained entirely within a single script/executable.
A single analysis driver with filters: Function evaluation is explicitly split into pre-processing (performed by the input filter), analysis, and post-processing (by the output filter).
A single analysis driver with environment variables: Function evaluation is contained within one analysis driver, but it requires environment variables to be set before running.
Multiple analysis drivers: Drivers are run sequentially or concurrently (See the asynchronous keyword) and can have any of the above options as well.

For fork and system interfaces, the analysis_driver list contains the names of one or more executable programs or scripts taking parameters files as input and producing results files as output. The first field in each analysis driver string must be an executable program or script for Dakota to spawn to perform the function evaluation. Drivers support:

One set of nested quotes, for arguments with spaces
Dakota will define special environment variables DAKOTA_PARAMETERS_FILE and DAKOTA_RESULTS_FILE which can be used in the driver script.
Variable definitions preceding the executable program or script, such as 'MY_VAR=2 run_analysis.sh' are no longer supported.

For details and examples see the Simulation Interface Components section of the Interfaces chapter of the User's Manual; for details on the filters and environment variables, see the subsection on Syntax for Filter and Driver Strings.

Examples

Examples:

  1. analysis_drivers = 'run_simulation_part1.sh' 'run_simulation_part2.sh'

  2. analysis_driver = 'run_simulation.sh -option "option 1"'

  3. analysis_driver = 'simulation.exe -option value -dakota_params $DAKOTA_PARAMETERS_FILE -input sim.in -dakota_results_file $DAKOTA_RESULTS_FILE'

FAQ

Where will Dakota look for the analysis_driver? Dakota will locate analysis_driver programs first in (or relative to) the present working directory (".", the interface-analysis_drivers-fork-work_directory if used, otherwise the directory in which Dakota is started), then the directory from which Dakota is started, then using the system $PATH environment variable (Path% on Windows).

Where should the driver be located? When the driver is a script it is most commonly placed in the same directory as the Dakota input file. When using a work_directory, Dakota will also look for drivers in the specified working directory, so link_files or copy_files may specify the driver to get copied or linked into the work directory. When executable programs are used as drivers, they are often elsewhere on the filesystem. These can be specified using absolute paths, or by prepending the PATH environment variable so Dakota finds them.

What if Dakota fails to run my analysis_driver? Prepend the absolute location of the driver to the PATH environment variable before running Dakota, or specify an absolute path to the driver in the Dakota input file.