Speeding up writing a very large text file

Question

davidwriter 2018-9-17

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/419329-speeding-up-writing-a-very-large-text-file

编辑： davidwriter 2018-9-21

I have to write very large text files (1-4 million lines; 100-400 MB) to disk. Writing binary output is NOT an option. To solve this problem I broke the output into blocks of 10000 lines and then used sprintf to write the formatted lines (each ending with a '\n') to a block of 10000 strings (dataop=string(10000,1)).

I have already opened the output file with the 'W' option and then successively write each block of strings using the command: fprintf(fid,dataop(:));

It is still taking an inordinate amount of time - each 1 million lines takes 60 minutes!

The machine is high-end: running Windows 10, 128GB RAM, dual Xeon 10 core processors, and WD Black 6TB drive. There are 5 other hard drives with are not active during the write process. I'm running 2017b

So - why is it so slow? Am I doing something stupid? (I'm more used to C++ than MATLAB)

2 个评论
显示无隐藏无

Stephen23 2018-9-17

@davidwriter: are the data numeric, strings, or char vectors? Do you really need to call sprintf as an intermediate step, why not just call fprintf directly?

davidwriter 2018-9-17

The data is numeric, but in any one line each of the elements has different format. I tried using fprintf but it was very slow, so I was trying to speed it up - after reading Yair Altman and Loren I used the sprintf and block approach to try to speed it up - in fact it made it worse.

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Stephen23 2018-9-17

3
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/419329-speeding-up-writing-a-very-large-text-file#answer_337127

编辑：Stephen23 2018-9-17

在 MATLAB Online 中打开

"To solve this problem I broke the output into blocks ..."

What problem? MATLAB has no problem with fprintf used on millions of elements. I don't see any reason, based on your explanation so far, why you need to process this in "blocks".

The problem is likely to be how you have coded this, but note that you forgot to actually upload your code, thus making it impossible for us to know what you are actually doing.

I just wrote a short script to test how long MATLAB requires to print 1 million character vectors (each char vector is arbitrarily sized 1x128) to a file:

C = cellstr(char(randi([33,126],1e6,128)));
tic
fid = fopen('temp4A.txt','wt');
fprintf(fid,'%s\n',C{:});
fclose(fid);
toc

It created a 127 megabyte file in

Elapsed time is 3.06831 seconds.

Then I tried a script with 16 floating point numbers per row, for 1 million rows:

N = 16;
M = rand(1e6,N);
F = repmat(',%.6f',1,N);
F = [F(2:end),'\n'];
tic
fid = fopen('temp4B.txt','wt');
fprintf(fid,F,M.');
fclose(fid);
toc

It created a 141 megabyte file in

Elapsed time is 28.429 seconds.

I can't get anywhere close to "each 1 million lines takes 60 minutes!"

3 个评论
显示 1更早的评论隐藏 1更早的评论

davidwriter 2018-9-19

编辑：davidwriter 2018-9-19

在 MATLAB Online 中打开

You are quite right in your comments. The problem turned out to be two fold: 1) In the test I was doing, the dimension of the string file that I was writing to was missing the second value, so instead of 10000,1 it was 10000,10000! (C++ vs MATLAB!!) 2) There was some problem with the system - trying your examples I got values 50x larger than you did. Restarting MATLAB didn't help, but rebooting Win 10 gave me values lower than yours.

I did a quick test of the various methods of writing 1000000 lines

1) writing the data to a string array (sa=strings(1000000,1);) using sprintf took 58.14 s

a) fprintf(fid,'%s\n', sa(:)) took 48.94 s - Total = 107.08 s
b) converting the strings to a cellstr and then writing
   sac=cellstr(sa);
   fprintf(fid,'%s\n', sac{:}) took 2.62 s - Total =60.75 s

2. Writing directly to disk using fprintf(fid,%7.1f \t %8.2f \t,...

opening the O/P file with 'wt' - Total = 61.40 s
opening the O/P file with 'Wt' - Total = 53.87 s

So clearly using a string buffer is unnecessary and the fastest is to use fprintf with a file opened with the 'W' option which prevents the buffer being flushed between writes.

Thank you for taking the time to help.

Jan 2018-9-19

@davidwriter: In "fprintf(fid,%7.1f \t %8.2f \t,..." the "..." might be the most interesting part. If you post some running code, it would be much easier to suggest improvements. It is not clear to me, what the inputs are. But converting numerical values to a string and again to a cell string is an indirection. See sprintfc or compose to create the required output format directly.

Creating a C-Mex function might be an option. But without knowing exactly what the inputs are, writing some code would include too much guessing.

davidwriter 2018-9-21

编辑：davidwriter 2018-9-21

在 MATLAB Online 中打开

Once again Stephen, my thanks, you are teaching me a lot.

The problem is fairly simple - I have ten arrays of dimension(nd, 1), where nd is usually between 1000000 - 4000000.

These arrays are: X Y Z Vx Vy Vz A R Pratio Frame all are stored as double, but some, like Frame, are integers and have to be written as such to the file. For the tests I created ten arrays dimensioned (1000000,1), filled with random numbers having the same range of values as the real data.

I tried to use sprintfc but it is very slow with complex formats. I used compose instead.

tstart=tic; 
dopc=compose('%8.1f \t %8.1f \t %7.2f \t %7.1f \t %7.1f \t %7.1f \t %6.2f \t %d \t %7.3f \t %d',...
       X, Y, Z, Vx, Vy, Vz, A, R, Pratio, Frame);
tst=toc(tstart);
disp(['Time to create string store ' num2str(tst)]);
avrfile = 'testcompose.avr';
disp(['Using compose : writing AVR file ' avrfile ' ...']);
fid_stm = fopen(stmfile,'Wt');
fprintf(fid_stm,'%s\n',dopc{:});
twstr=toc(tstart);
disp(['Compose: Writing time = ' num2str(twstr-tst) ' - Total time = ' num2str(twstr)]);

which gave the result:

Time to create string store: 279.9028
Using compose : writing AVR file testcompose.avr ...
Compose: Writing time = 2.0855 - Total time = 281.9884

Not good - but if I replaced the second line above with the following:

dataf = [X Y Z Vx Vy Vz A R Pratio Frame];
dopc=compose('%8.1f \t %8.1f \t %7.2f \t %7.1f \t %7.1f \t %7.1f \t %6.2f \t %d \t %7.3f \t %d',...
         dataf);

A dramatic improvement -

Time to create string store: 11.691
Using compose : writing AVR file testcompose.avr ...
Compose: Writing time = 2.0849 - Total time = 13.8387

So it seems that this latter approach is the fastest and the easiest - 5x faster than direct fprintf with the 'W' option.

请先登录，再进行评论。

Answer 2

KSSV 2018-9-17

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/419329-speeding-up-writing-a-very-large-text-file#answer_337113

在 MATLAB Online 中打开

How about this approach?

S = rand(10000,3) ;
S = cellstr(num2str(S)) ;
fid = fopen('data.txt','w') ;
fprintf(fid,'%s\n',S{:});
fclose(fid) ;

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

Speeding up writing a very large text file

2 个评论
显示无隐藏无

采纳的回答

3 个评论
显示 1更早的评论隐藏 1更早的评论

更多回答（1 个）

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

另请参阅

类别

标签

产品

Community Treasure Hunt

Speeding up writing a very large text file

2 个评论 显示 无隐藏 无

采纳的回答

3 个评论 显示 1更早的评论隐藏 1更早的评论

更多回答（1 个）

0 个评论 显示 -2更早的评论隐藏 -2更早的评论

另请参阅

类别

标签

产品

Community Treasure Hunt

2 个评论
显示无隐藏无

3 个评论
显示 1更早的评论隐藏 1更早的评论

0 个评论
显示 -2更早的评论隐藏 -2更早的评论