How would parallel computing speed up the analysis?

Tags: tutorial qsub peer distcomp matlab meg-language

Introduction

Many times you are faced with the analysis of multiple subjects and experimental conditions, or with the analysis of your data using multiple analysis parameters (e.g., frequency bands). Parallel computing in MATLAB can help you to speed up these types of analysis.

Note that this is usually referred to as distributed computing if you are submitting multiple independent computations to multiple computers. The term parallel computing is usually reserved for multiple CPUs or computers working simultaneously at the same problem that requires constant sharing of small snippets of data between the CPUs. Since the analyses of multiple subjects are done independently of each other, we call it distributed computing.

This tutorial describes two approaches for distributing the analysis of multiple subjects and conditions. The data used in this example is the same as in the tutorial scripts: 151-channel MEG was recorded in 4 subjects, and in each dataset there are three experimental conditions (FC, FIC, IC). Both approaches rely on the qsubcellfun function which applies a given function to each element of a cell-array. The function execution is done in parallel on the Torque batch queue system.

After this tutorial you should be able to execute your multi-subject analysis in parallel and design analysis scripts that allow for easy parallelization, either over subjects or over parameters used in the analysis.

In this tutorial we use the qsub toolbox that is released along with FieldTrip. There are alternative methods for distributed computing, such as the MATLAB Parallel Computing toolbox (e.g., using parfor or batch) or with the peer-to-peer toolbox (also included with FieldTrip). More general information about the different approaches for distributed processing in MATLAB can be found in the frequently asked questions.

Background

The qsub distributed computing toolbox was implemented with FieldTrip (and SPM) in mind. At the moment however, FieldTrip itself does not yet make use of distributed computing, i.e. FieldTrip functions do not automatically distribute the workload. We are of course planning to make that possible, i.e. that a single cfg.parallel=’yes’ option will automatically distribute the computational load over all available nodes.

At the moment the only way of distributing the workload over multiple nodes requires that you adapt your scripts. The easiest is to distribute the workload of the analysis of multiple subjects and conditions over multiple nodes.

Procedure

To distribute your processes and to speed up your analyses, we provide two examples. The first example script will show you how to use basic FieldTrip functions for the distribution. Using the basic FieldTrip functions in a memory efficient manner requires that you save the intermediate data of each step to disk, and that you load it upon the next (parallel) step in the analysis.

If you prefer not to store all intermediate results, or if you want to have more control over other aspects of the parallel execution, you can provide your own functions that are executed in parallel. This is demonstrated in the second example script.

The distributed operations of FieldTrip functions in this example require the original MEG datasets for the four subjects, which are available from

• Subject01.zip
• Subject02.zip
• Subject03.zip
• Subject04.zip

Example 1: using only FieldTrip functions in distributed computing

This example script demonstrates how to run basic FieldTrip functions in parallel. The idea is schematically depicted in the following figure.

subjectlist = {
  'Subject01.ds'
  'Subject02.ds'
  'Subject03.ds'
  'Subject04.ds'
  };

conditionlist = {
  'FC'
  'FIC'
  'IC'
  };

triggercode = [
  9
  3
  5
  ];

% start with a new and empty configuration
cfg = {};

for subj=1:4
  for cond=1:3
    cfg{subj,cond}                      = [];
    cfg{subj,cond}.dataset              = subjectlist{subj};
    cfg{subj,cond}.trialdef.prestim     = 1;
    cfg{subj,cond}.trialdef.poststim    = 2;
    cfg{subj,cond}.trialdef.eventtype   = 'backpanel trigger';
    cfg{subj,cond}.trialdef.eventvalue  = triggercode(cond);
  end
end

cfg = qsubcellfun(@ft_definetrial, cfg, 'memreq', 4*1024^3, 'timreq', 10*60);

% this extends the previous configuration
for subj=1:4
  for cond=1:3
    cfg{subj,cond}.channel              = {'MEG', '-MLP31', '-MLO12'};
    cfg{subj,cond}.demean               = 'yes';
    cfg{subj,cond}.baselinewindow       = [-0.2 0];
    cfg{subj,cond}.lpfilter             = 'yes';
    cfg{subj,cond}.lpfreq               = 35;
  end
end

data = qsubcellfun(@ft_preprocessing, cfg, 'memreq', 4*1024^3, 'timreq', 10*60);

% start with a new and empty configuration
cfg = {};

for subj=1:4
  for cond=1:3
    % timelockanalysis does not require any non-default settings
    cfg{subj,cond} = [];
  end
end

timelock = qsubcellfun(@ft_timelockanalysis, cfg, data, 'memreq', 4*1024^3, 'timreq', 10*60);

% from here on we won't process the data in parallel any more
% average each condition over all subjects
cfg = [];
avgFC  = ft_timelockgrandaverage(cfg, timelock{:,1});
avgFIC = ft_timelockgrandaverage(cfg, timelock{:,2});
avgIC  = ft_timelockgrandaverage(cfg, timelock{:,3});

cfg = [];
cfg.layout = 'CTF151_helmet.mat';
ft_multiplotER(cfg, avgFC, avgFIC, avgIC);

cfg = [];
cfg.channel = {'MLC33', 'MLC43', 'MLP11', 'MLP12', 'MLP13', 'MLP33', 'MLP34', 'MLT14', 'MLT15', 'MLT25'}
ft_singleplotER(cfg, avgFC, avgFIC, avgIC);

In the code above all data is processed by the distributed computers and subsequently returned to the workspace of your desktop computer. The data can take quite a lot of RAM, which you can check like this.

>> whos
  Name               Size                 Bytes  Class     Attributes

  conditionlist      3x1                    350  cell
  subjectlist        4x1                    544  cell
  triggercode        3x1                     24  double
  cfg                4x3                9732384  cell
  data               4x3             1203237420  cell
  timelock           4x3               65515200  cell
  ...

Instead of returning the 12 variables for the different subjects and conditions all to your workspace, you can also use the cfg.inputfile and cfg.outputfile options to have the distributed computers read/write the data to/from disk. For example the section on ft_preprocessing and ft_timelockanalysis could be changed into

% ...

for subj=1:4
  for cond=1:3
    cfg{subj,cond}.channel              = {'MEG', '-MLP31', '-MLO12'};
    cfg{subj,cond}.demean               = 'yes';
    cfg{subj,cond}.baselinewindow       = [-0.2 0];
    cfg{subj,cond}.lpfilter             = 'yes';
    cfg{subj,cond}.lpfreq               = 35;
    cfg{subj,cond}.outputfile           = sprintf('subj%02d_cond%02d_raw.mat', subj, cond);
  end
end

% note that here we don't specify an output parameter
qsubcellfun(@ft_preprocessing, cfg);

cfg = {};

for subj=1:4
  for cond=1:3
    cfg{subj,cond}.inputfile  = sprintf('subj%02d_cond%02d_raw.mat', subj, cond);
    cfg{subj,cond}.outputfile = sprintf('subj%02d_cond%02d_avg.mat', subj, cond);
  end
end

% note that here we don't specify an output parameter
qsubcellfun(@ft_timelockanalysis, cfg);

% ...

Example 2: writing custom functions for distributed computing

This example script demonstrates how you can efficiently design your custom code for distributed computing.

subjectlist = {
  'Subject01.ds'
  'Subject02.ds'
  'Subject03.ds'
  'Subject04.ds'
  };

conditionlist = {
  'FC'
  'FIC'
  'IC'
  };

triggercode = [
  9
  3
  5
  ];

% start with a new and empty configuration
cfg1 = {};
cfg2 = {};
cfg3 = {};
cfg4 = {};
outputfile = {};

for subj=1:4
  for cond=1:3
    % this is for definetrial and preprocessing
    cfg1{subj,cond}                      = [];
    cfg1{subj,cond}.dataset              = subjectlist{subj};
    cfg1{subj,cond}.trialdef.prestim     = 1;
    cfg1{subj,cond}.trialdef.poststim    = 2;
    cfg1{subj,cond}.trialdef.eventtype   = 'backpanel trigger';
    cfg1{subj,cond}.trialdef.eventvalue  = triggercode(cond);
    cfg1{subj,cond}.channel              = {'MEG', '-MLP31', '-MLO12'};
    cfg1{subj,cond}.demean               = 'yes';
    cfg1{subj,cond}.baselinewindow       = [-0.2 0];
    cfg1{subj,cond}.lpfilter             = 'yes';
    cfg1{subj,cond}.lpfreq               = 35;

    % this is for timelockanalysis
    cfg2{subj,cond}                      = [];

    % this is for megplanar
    cfg3{subj,cond}                      = [];

    % this is for combineplanar
    cfg4{subj,cond}                      = [];

    % this defines the file that will contain the output
    outputfile{subj,cond} = sprintf('subj%02d_cond%02d_combined.mat', subj, cond);
  end
end

% note that the "preproc_timelock_planar" function is defined further down in this tutorial
qsubcellfun(@preproc_timelock_planar, cfg1, cfg2, cfg3, cfg4, outputfile);

% let's now load the individual subject data from the 12 .mat files and average each of them for subsequent plotting
timelock = {};
for subj=1:4
  for cond=1:3
    tmp = load(sprintf('subj%02d_cond%02d_combined.mat', subj, cond));
    timelock{subj,cond} = tmp.combined;
    clear tmp
  end
end

cfg = [];
avgFC  = ft_timelockanalysis(cfg, timelock{:,1));
avgFIC = ft_timelockanalysis(cfg, timelock{:,2));
avgIC  = ft_timelockanalysis(cfg, timelock{:,3));

cfg = [];
cfg.layout = 'CTF151_helmet.mat';
ft_multiplotER(cfg, avgFC, avgFIC, avgIC)

This way you can distribute your custom function (e.g., see below) along with the input and output parameters.

function preproc_timelock_planar(cfg1, cfg2, cfg3, cfg4, outputfile)

cfg1 = ft_definetrial(cfg1);
data = ft_preprocessing(cfg1);

timelock = ft_timelockanalysis(cfg2, data);
clear data

planar = ft_megplanar(cfg3, timelock);
clear timelock

combined = ft_combineplanar(cfg4, planar);
clear planar

save(outputfile, 'combined');
clear combined

Summary and suggested further reading

This tutorial covered how to distribute your computations/workload over multiple computers in a cluster that uses the Torque or SGE batch queue system. In our example, we have performed a relatively simple timelock analysis to compute event-related fields, but one can imagine that it does not need many adjustments to distribute any other type of analysis. Using the configuration demonstrated in Example 2, you can distribute any form of analysis.

FAQs related to issues in this tutorial:

• How to compile MATLAB code into stand-alone executables?
• How to get started with distributed computing using qsub?
• What are the different approaches I can take for distributed computing?

• Does a firewall affect the communication between peers?
• How can I combine FieldTrip with peer distributed computing?
• How can I debug a problematic distributed job?
• How can I determine the number of threads that MATLAB uses?
• How can I prevent a job from executing twice?
• How can I read and write files if I use other people's peers?
• How can I set up the peer distributed computing on a large Linux cluster?
• How can I set up the peer distributed computing on a single multicore computer?
• How can I set up the peer distributed computing on a small number of computers?
• How can I stop the different threads created by peercontroller and peerworker?
• How can I use the command-line peerworker and optimize the MATLAB licenses?
• How do I avoid having to allocate N copies of my data if I want to execute N jobs?
• How does the smartmem algorithm work?
• How does the peer smartshare algorithm work?
• How should I call peercellfun when a function requires many inputs (e.g., key-value pairs)?
• How to get started with peer distributed computing on my own desktop computer?
• What are the different approaches I can take for distributed computing?
• What happens if a peer distributed job fails to execute properly?
• What happens with a peer distributed job that has an error on the worker?
• Why are the peers using multicast to announce themselves?
• Why does peercellfun resubmit jobs that take too long to get started?

• How to get started with distributed computing using engine?
• What are the different approaches I can take for distributed computing?

• How to get started with the MATLAB distributed computing toolbox?
• What are the different approaches I can take for distributed computing?

What is speed up in parallel algorithm?

What is speedup and efficiency in parallel computing?

Why will the speedup of a parallel algorithm reach a limit?

How does parallel computing save time?