batch

Perform evaluations in batches

Topics

concurrency_and_parallelism

Specification

• Alias: None

• Arguments: None

• Default: sequential interface usage

Child Keywords:

Required/Optional	Description of Group	Dakota Keyword	Dakota Keyword Description
Optional		size	Limit the number of evaluations in a batch

Description

When the optional batch keyword is used, Dakota performs evaluations in batches. In batch mode, Dakota writes the parameters for multiple (a batch of) evaluations to a single file, invokes the analysis driver once for the whole batch, and expects to find results for the entire batch in a single file after the analysis driver has exited. Batch mode may be useful when a user desires to take greater control over job management. For example, the analysis driver can perform the evaluations in the batch in any sequence or in concurrent sub-batches.

The names of the parameters file and results file are provided as command line arguments to the analysis driver, just as in a conventional, non-batch evaluation. By default, all currently available evaluations will be performed in a single batch, but the batch size can be limited using the size keyword.

Batch mode has a few important limitations.

• Only one batch at a time may be executed. Asynchronous execution of multiple concurrent batches is not supported.

• No input_filter or output_filter is permitted.

• Only one analysis_driver is allowed.

• failure_capture modes are restricted to abort and recover.

Some of these restrictions may be lifted in future Dakota releases.

File Formats

A batch parameters file written by Dakota is simply a concatenated set of parameters files for the set of evaluations, either in aprepro or default Dakota format.

The batch results file is also a concatenated set of results files for the individual evaluations. However, because Dakota’s results file format is not as rich as its parameters file format, evaluations in the batch results file must be separated by a line that begins with the ‘#’ character.

The order of evaluations in the batch results file must match the order in the batch parameters file.

Tagging and Work Directories

When Dakota’s work_directory feature is enabled, one directory is created per batch. If file_tag or directory_tag is used, parameters/results files and work directories are tagged with a batch Id, which is an integer that begins with 1 and increments for each new batch.