Parallelize cumulative average algorithm

Hi,

is there a way how to parallelize algorithm computing average of data accessible only as a stream using parfor?

dat = 1:9;

avg = 0;

for i=1:numel(dat)

avg = (dat(i)+(i-1)*avg)/i;

end

avg

In reality the code looks more like the following, where the input data is too big to be stored.

avg = 0;

for i=1:numFrames

frm = getNewFrame(); % returns new data

avg = (frm+(i-1)*avg)/i;

end

I somehow understand why this code can not be parallelized as is, but maybe Matlab offers some trick.

Answer 1

[Edric Ellis]

You can use a parfor "reduction variable" to compute this, with a slight modification which may or may not be significant:

dat = rand(1,9);
avg = 0;
N = numel(dat);
parfor i = 1:N
    avg = avg + dat(i)/N;
end

Starting parallel pool (parpool) using the 'Processes' profile ...
Parallel pool using the 'Processes' profile is shutting down.

avg, mean(dat)

avg = 0.6135

ans = 0.6135