dlmwrite vs. writematrix: speed

dlmwrite, according to its page, is not recommended. It is suggested to use writematrix instead (better performance). I currently use dlmwrite to continuously append a text file with measurement data. Speed is therefore important. Comparing dlmwrite to writematrix,

fwrtmat = 'test writematrix.txt';
fdlmwrt = 'test dlmwrite.txt';

ncols = 10;                                     % number of columns
niter = 1e3;                                    % number of iterations in test
fmt = ['%f\t\n',repmat('%f\t ', 1, ncols)];     % format for dlmwrite
fmt(end:end+1) = '\n';
fid = fopen(fdlmwrt, 'w');
fclose(fid);

tic
for i = 1:niter
    data = rand(1,ncols);
    writematrix(data,fwrtmat,'WriteMode','append','FileType','text','Delimiter','tab');
end
toc

Elapsed time is 1.890530 seconds.

tic
for i = 1:niter
    data = rand(1,ncols);
    dlmwrite(fdlmwrt,data,'-append','delimiter','\t','precision',15);   % use same precision as writematrix
end
toc

Elapsed time is 0.283388 seconds.

I found that dlmwrite is faster, which I did not expect. Additionally, the (default?) precision of writematrix is not required and it would even be preferential to sacrifice precision for performance. The file format can be changed if that would improve performance.

Am I using writematrix correctly/optimally in my example? Should I use writematrix or stick to dlmwrite?

Answer 1

[Sulaymon Eshkabilov]

If you run this simulation on the MATLAB desktop and shift the orders, their elapsed time will be similar. Anyhow it is interesting :)

Here is another discussion thread demonstrating your conclusion, i.e., dlmwrite() is faster than writematrix(): https://www.mathworks.com/matlabcentral/answers/653363-fastest-way-to-append-file-with-matrix-data

fwrtmat = 'test writematrix.txt';
fdlmwrt = 'test dlmwrite.txt';

ncols = 10;                                     % number of columns
niter = 1e3;                                    % number of iterations in test
fmt = ['%f\t\n',repmat('%f\t ', 1, ncols)];     % format for dlmwrite
fmt(end:end+1) = '\n';
fid = fopen(fdlmwrt, 'w');
fclose(fid);

tic
for i = 1:niter
    data = rand(1,ncols);
    dlmwrite(fdlmwrt,data,'-append','Delimiter','\t','precision',15);   % use same precision as writematrix
end
TDLM=toc
tic
for i = 1:niter
    data = rand(1,ncols);
    writematrix(data,fwrtmat,'WriteMode','append','FileType','text','Delimiter','\t');
end
TWRTM=toc

[tetra]

Thank you for recommending that thread. It appears I did not search Matlab answers thoroughly. I explored the suggestion of using a binary file. Using similar code:

fbinwrt = 'test binwrite.bin';
ncols = 10;                                     % number of columns
niter = 1e3;                                    % number of iterations
fidbin = fopen(fbinwrt,'a');                    % open binary file, 'a' = append data

tic
for i = 1:niter
    data = rand(1,ncols);
    fwrite(fidbin,data,'double');               % format double
end
fclose(fidbin);                                 % close before reading
TBIN = toc

TBIN = 0.0063

This is ~100 times as fast as dlmwrite on desktop!

Now to open the file (less straightforward to me than when using dlmwrite and writematrix):

bincheck=fopen(fbinwrt,'r');                    % 'r' = read data
bindata=fread(bincheck,[ncols,inf],'double')';  % transposing will result in 'correct' size
fclose(bincheck);

And indeed as you mentioned, the elapsed times between dlmwrite and writematrix were somewhat similar on my Matlab desktop. dlmwrite was about twice as fast as writematrix.