Batch Job

A MATLAB toolbox to parallelize simple for loops across multiple MATLAB instances, across multiple computing nodes. For the toolbox to work, its root directory needs to be on your MATLAB path at startup.

Overview

This toolbox can parallelize for loops which are of the form:

for a = 1:size(input, 2)
    output(:,a) = func(input(:,a), global_data);
end

where input and output can be numeric or cell arrays of any shape or size, and global_data can be anything. The for loop iterates over the last non-singleton dimension of input, and output is concatenated along the first singleton dimension of a single function output.

The for loop is parallelized across multiple MATLAB instances which may or may not be on the same computer, depending on which toolbox function is used.

The for loop is straightforwardly replaced with one or more function calls to batch_job* functions, such as:

output = batch_job_distrib(func, input, {'', num_workers}, global_data);

here, num_workers being the number of local MATLAB instances to parallelize over.

The functionality in this toolbox essentially replicates that of parfor, without the need for the Parallel Computing Toolbox or a Distributed Computing Server.

In addition, it has some other benefits:

• Errors are caught and the error message stored, but the for loop continues.
• The -progress option shows a progress bar for the loop.
• The -timeout option allows a timeout to be specified, which limits each iteration to a maximum allowed computation time.
• The -async option allows the loop computation to be done in parallel to other computation in the main thread (batch_job_distrib() only).

Three approaches

The toolbox provides three approaches for parallelizing for loops:

1. Single call function, batch_job_distrib(), which spawns MATLAB instances on the specified computers, and uses the file system to communicate between workers.
2. Low level functions, batch_job_submit() and batch_job_collect(), which spread work across worker MATLAB instances (across multiple PCs) which are running batch_job_worker(), using the file system to communicate.
3. Single call function, batch_job(), which spawns MATLAB instances locally, and uses a memory mapped file to communicate between workers. This approach does not handle heterogeneous (non-uniform) outputs, nor does support asynchronous computation, but is able to kill more hung processes than batch_job_distrib() using the -timeout option.

See the help text for each of those functions for usage.

Reporting bugs

This toolbox has a lot of functionality, and it's difficult to make sure it all works in all scenarios. If you find it's not working for you, please run batch_job_test() to make sure that works as expected (i.e. doesn't report any errors). If it does, feel free to report an issue. If it fails, please make sure that the batch_job folder is on the path at startup for all workers.