perl text mining code can't handle massive amounts of data - performance

Question

I'm doing a large text mining project. I have 100,000 text files. I've extracted two- and three-word phrases from sets of 1,000 documents at a time and have created 100 files. Each file has roughly 8 million lines in this format:
total_references num_docs_referencing_phrase phrase
I want to create an aggregate list of total references and number of docs referencing each phrase by processing the 100 intermediate files. To that end I wrote this program.
#!/usr/bin/perl -w
$| = 1 ; # Don't buffer output
use File::Find ;
$dir = "/home/sl/phrase-counts" ;
find(\&processFile, $dir) ;
for $key ( keys %TOTALREFS ) {
print "$TOTALREFS{$key} $NUMDOCS{$key} ${key}\n" ;
}
sub processFile {
my $file = $_ ;
my $fullName = $File::Find::name ;
if ( $fullName =~ /\.txt$/ ) {
$date = `date` ;
chomp $date ;
print "($date) file: $fullName\n" ;
open INFILE, "$fullName" or die "Cannot read ${fullName}";
while ( <INFILE> ) {
my $line = $_ ;
chomp $line ;
( $totalRefs, $numDocs, $phrase ) = split (/\s+/, $line, 3) ;
$TOTALREFS{$phrase} += $totalRefs ;
$NUMDOCS{$phrase} += $numDocs ;
}
close ( INFILE ) ;
}
}
The code produces strange errors after 8 or so files are processed and then it hangs, i.e. it stops listing files it should be processing.
Use of uninitialized value $date in scalar chomp at ./getCounts line 21.
Use of uninitialized value $date in concatenation (.) or string at ./getCounts line 22.
I don't believe the problem is really my date command, especially since it runs fine for a number of early files processed and because the problem does not occur at the same point in the run every time I run it. I assume the problem is that my program is consuming too much system resource and corrupting the state of the running environment. Running top and watching memory use go up to 97% of the machine concerns me although I notice that the errors and hang occur before top shows little memory left. And, there is some swap on the machine.
My question is, how can I rewrite this program to actually complete its execution? With 8 million lines of data for each of 100 files there could be 800 million lines of output although I would guess that the total is more likely in the range of 50-100 million lines. I have done some cleanup of the data and could consider more aggressive sanitizing of phrases to cut down on the numbers but I'd like to understand how I can design this code better.
I've seen articles that tell programmers to put their data into a database. My concern is the time it might take to update a database 100 million times.
Suggestions?

Answer 1

It looks like you're running on a *nix system, so make sort do all the work for you. It knows how to use memory efficiently.
sort -k 3 all_your_input_files*.txt > sorted.txt
Why do this? Because now all lines corresponding to the same phrase appear in a single block within the file, so you can compute totals easily: just write a short Perl script that adds the current line's numbers to the current totals, and writes them out whenever the phrase changes from the previous line (and at the end):
my ($oldPhrase, $totTotalRefs, $totNumDocs) = (undef, 0, 0);
while ( <INFILE> ) {
my $line = $_ ;
chomp $line ;
( $totalRefs, $numDocs, $phrase ) = split (/\s+/, $line, 3) ;
if (defined($oldPhrase) && $phrase ne $oldPhrase) {
print "$totTotalRefs $totNumDocs $oldPhrase\n" ;
$totTotalRefs = $totNumDocs = 0;
}
$totTotalRefs += $totalRefs ;
$totNumDocs += $numDocs ;
$oldPhrase = $phrase;
}
close ( INFILE ) ;
print "$totTotalRefs $totNumDocs $oldPhrase\n" ;
The above code is untested, but should work with appropriate boilerplate added I think.
[EDIT: Fix bug in which $oldPhrase never gets set, as suggested by Sol.]

Answer 2

You are storing all of the different phrases as keys for both %TOTALREFS and %NUMDOCS, so things are at least twice as bad as they need to be.
I suggest you try the following
Add use strict and use warnings (instead of -w) and declare all of your variables properly
Don't use capitals in your variable names. Capital letters are reserved for global identifiers
Don't start 100 subprocesses just to get the time of day. Just use localtime like this
printf "(%s) file: %s\n", scalar localtime, $full_name;
Use find just to generate an array of the files to be processed, so it would look like this
my #files;
find(sub {
push #files, $File::Find::name if -f and /\.txt$/i;
}, $dir) ;
Then you can process each file with a simple for loop
for my $file (#files) {
...
}
Take two passes through the files, the first time generating a hash that relates each phrase to an integer starting at zero, and the second that uses those integers to index arrays #total_refs and #num_docs and increment their elements
You may still run out of memory, but those measures will certainly give you a better chance.
Update
Just to be clear, this is how I imagine it would work. I've done this as a single pass, but it may be better to write it as two passes as I described so that you can check your intermediate data.
Note that this isn't tested apart from making sure that it compiles.
#!/usr/bin/perl
use strict;
use warnings;
use 5.010;
use autodie;
STDOUT->autoflush;
use File::Find;
my $dir = '/home/sl/phrase-counts';
my #files;
find(sub {
push #files, $File::Find::name if -f and /\.txt$/i;
}, $dir);
my (%phrases, #total_refs, #num_docs);
my $num_phrases = 0;
for my $file (#files) {
printf "(%s) file: %s\n", scalar localtime, $file;
open my $in_fh, '<', $file;
while (<$in_fh>) {
chomp;
my ($total_refs, $num_docs, $phrase) = split ' ', $_, 3;
my $phrase_num = $phrases{$phrase} //= $num_phrases++;
$total_refs[$phrase_num] += $total_refs;
$num_docs[$phrase_num] += $num_docs;
}
}
for my $phrase (keys %phrases) {
my $phrase_num = $phrases{$phrase};
printf "%s %s %s\n",
$total_refs[$phrase_num],
$num_docs[$phrase_num],
$phrase_num;
}

Answer 3

Trying to use more resources than available causes exceptions for being unable to allocate memory or results in system calls returning error messages. It doesn't corruption memory.
In this case, the result of backticks is undef, which means the command could not be executed. That could very well be because you have insufficient memory left. Where did you get the idea that being unable to execute a program is the result of corrupted memory?! Furthermore, you have an error you don't understand, yet you didn't check what error was returned? Backticks sets $? (and $! when $? is negative) as per system. Assuming it's a bug in Perl is a very bad assumption to make, especially when the system tells you what error occurred.
Use less memory, either through the use of a more appropriate and/or efficient data structure, or by keeping a portion of the data out of memory (e.g. on disk or in a database).