Can I read a gigantic text file with Parallel Computing? - matlab

Question

I have multiple text files that are about 2GB in size (approximately 70 million lines). I also have a quad-core machine and access to the Parallel Computing toolbox.
Typically you might open a file and read lines as so:
f = fopen('file.txt');
l = fgets(f);
while ~ isempty(l)
% do something with l
l = fgets(f);
end
I wanted to distribute the "do something with l" across my 4 cores, but that of course requires the use of a parfor loop. That would require that I "slurp" the 2GB file (to borrow a Perl term) into MATLAB a priori, instead of processing on the fly. I don't actually need l, just the result of the processing.
Is there a way to read lines out of a text file with parallel computing?
EDIT: It's worth mentioning that I can find the exact number of lines ahead of time (!wc -l mygiantfile.txt).
EDIT2: The structure of the file is as follows:
15 1180 62444 e0e0 049c f3ec 104
So 3 decimal numbers, 3 hex numbers, and 1 decimal number. Repeat this for 70 million lines.

Answer 1

As requested, I'm showing an example of memory-mapped files using memmapfile class.
Since you didn't provide the exact format of the data file, I will create my own. The data I am creating is a table of N rows, each consisting of 4 columns:
first is a double scalar value
second is a single value
third is a fixed-length string representing a uint32 in HEX notation (e.g: D091BB44)
fourth column is a uint8 value
The code to generate the random data, and write it to binary file structured as described above:
% random data
N = 10;
data = [...
num2cell(rand(N,1)), ...
num2cell(rand(N,1,'single')), ...
cellstr(dec2hex(randi(intmax('uint32'), [N,1]),8)), ...
num2cell(randi([0 255], [N,1], 'uint8')) ...
];
% write to binary file
fid = fopen('file.bin', 'wb');
for i=1:N
fwrite(fid, data{i,1}, 'double');
fwrite(fid, data{i,2}, 'single');
fwrite(fid, data{i,3}, 'char');
fwrite(fid, data{i,4}, 'uint8');
end
fclose(fid);
Here is the resulting file viewed in a HEX editor:
we can confirm the first record (note that my system uses Little-endian byte ordering):
>> num2hex(data{1,1})
ans =
3fd4d780d56f2ca6
>> num2hex(data{1,2})
ans =
3ddd473e
>> arrayfun(#dec2hex, double(data{1,3}), 'UniformOutput',false)
ans =
'46' '35' '36' '32' '37' '35' '32' '46'
>> dec2hex(data{1,4})
ans =
C0
Next we open the file using memory-mapping:
m = memmapfile('file.bin', 'Offset',0, 'Repeat',Inf, 'Writable',false, ...
'Format',{
'double', [1 1], 'd';
'single', [1 1], 's';
'uint8' , [1 8], 'h'; % since it doesnt directly support char
'uint8' , [1 1], 'i'});
Now we can access the records as an ordinary structure array:
>> rec = m.Data; % 10x1 struct array
>> rec(1) % same as: data(1,:)
ans =
d: 0.3257
s: 0.1080
h: [70 53 54 50 55 53 50 70]
i: 192
>> rec(4).d % same as: data{4,1}
ans =
0.5799
>> char(rec(10).h) % same as: data{10,3}
ans =
2B2F493F
The benefit is that for large data files, is that you can restrict the mapping "viewing window" to a small subset of the records, and move this view along the file:
% read the records two at-a-time
numRec = 10; % total number of records
lenRec = 8*1 + 4*1 + 1*8 + 1*1; % length of each record in bytes
numRecPerView = 2; % how many records in a viewing window
m.Repeat = numRecPerView;
for i=1:(numRec/numRecPerView)
% move the window along the file
m.Offset = (i-1) * numRecPerView*lenRec;
% read the two records in this window:
%for j=1:numRecPerView, m.Data(j), end
m.Data(1)
m.Data(2)
end

Answer 2

Some matlab's built-in functions support multithreading - the list is here. There is no need for the Parallel Computing toolbox.
If the "do something with l" can benefit from the toolbox, just implement the function before reading another line.
You may alternatively want to read the whole file using
fid = fopen('textfile.txt');
C = textscan(fid,'%s','delimiter','\n');
fclose(fid);
and then compute the cells in C in parallel.
If the reading time is a key issue, you may also want to access parts of the data file within a parfor loop. Here is an example from Edric M Ellis.
%Some data
x = rand(1000, 10);
fh = fopen( 'tmp.bin', 'wb' );
fwrite( fh, x, 'double' );
fclose( fh );
% Read the data
y = zeros(1000, 10);
parfor ii = 1:10
fh = fopen( 'tmp.bin', 'rb' );
% Get to the correct spot in the file:
offset_bytes = (ii-1) * 1000 * 8; % 8 bytes/double
fseek( fh, offset_bytes, 'bof' );
% read a column
y(:,ii) = fread( fh, 1000, 'double' );
fclose( fh );
end
% Check
assert( isequal( x, y ) );