Background Data Generation

Workflow Overview

Running simulations without interruption requires a supply of input data. For example, you might want to:

In this example, you compare the run times for producing input data and processing data frames in a conventional simulation for-loop versus a background runner simulation for-loop. The helperBackgroundRunner supporting file implements an object that enables you to generate simulation input data in the background while your simulation loop runs on parameterized conditions.

The helperBackgroundRunner object runs your data‐generation function on a pool of background workers by using the backgroundPool and parfeval function and buffers the results in a FIFO queue. You specify the generator function and the desired number of outputs as input arguments to the helperBackgroundRunner object. The object creates the pool, schedules futures, and returns the generated data. For more details on the underlying parallel mechanisms, see the parfeval function and Background Workers.

This examples uses a simple workflow to demonstrate how the helperBackgroundRunner object can provide real-time access to simulation input data. It uses:

Create Generator Object

Define a generator that produces one frame of channel estimates and a background runner object.

genFcn = @() genChanout();
br = helperBackgroundRunner( ...
    GeneratorFunction=genFcn, ...
    NumOutputs=2);

When you create a background runner object to generate data:

Process Frames in Conventional Loop

Simulate frame processing in a conventional for-loop, and then display the total run time.

numIterations = 1000;
tic
for k = 1:numIterations
[rxSignalFM, pathGainsFM] = genChanout();
    % Next, a practical simulation would recover the received
    % waveform and compare to the transmitted waveform to
    % compute a BER.
end
frameModeTime = seconds(toc)

frameModeTime = duration
   38.015 sec

Process Frames in Background Loop

Replace direct calls to your batch generator with a single call to the helperBackgroundRunner object, br. Each call fetches the next ready batch. The helperBackgroundRunner automatically schedules a new future to regenerate that slot in the background. Check the number of workers on your system by assigning the variable pool.

pool = backgroundPool

pool = 
 BackgroundPool with properties: 

           NumWorkers: 6
                 Busy: false

tic
for ii = 1:numIterations
    % Pull the next ready batch, blocking only if the queue is empty
[rxSignal, pathGains] = br();
    % Next, a practical simulation would recover the received
    % waveform and compare to the transmitted waveform to
    % compute a BER.
end
brTime = seconds(toc)

brTime = duration
   14.327 sec

Startup time for the parallel pool will result in increased time for first simulation. Subsequent runs after the parallel pool is initialized will show true benefit of running with background pool. Your training loop never stalls waiting for data because the helperBackgroundRunner object maintains its own queue and spawns futures on each fetch.

Inspect Results

When the simulation ends, display the number of workers and futures that the br object used.

s = info(br); % Display worker and futures status
fprintf("helperBackgroundRunner using %d workers, %d futures.\n", ...
s.NumWorkers,s.NumFutures);

helperBackgroundRunner using 6 workers, 7 futures.

Release Background Jobs

Next, terminate all background jobs by releasing the br object.

release(br);  % Terminate all background jobs

Local Function

function [rxSignal, pathGains] = genChanout()
    waveform = wwg_qam();
    [rxSignal, pathGains] = wcd_mimo(waveform);
end