Optimization anyone

ruby-talk
1

I have this little script that takes a list of keyword sets, each set has only
two keywords and for each one of them the script creates a regular expression
like this:

Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")

then I match it to a string that contains a long text fetched from a web page.

a more complete pseudo-code

···

#########################################
long_text = get_web_page(url)

keyword_hash = load_keyword_array_from_database

keyword_hash.each_pair { |id, value|
  
    key1 = value[0]
    key2 = value[1]
      
    r = Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")
    return id if long_text =~ r
}

return -1
###########################################

Now this code works perfect, the problem is that the keyword_hash has more
than 300 elements and running this code can take between 50 to 120 seconds.
Since I am processing more than 1000 pages with this code it takes forever.

I solved this problem by replacing the regular expression match to

   r1 = Regexp.new("#{key1}\.*#{key2}")
   r2 = Regexp.new("#{key2}\.*#{key1}")
   
   return id if long_text =~ r1 or long_text =~ r2

I simply put the or statement outside the regular expresion and the speedup
was from 50~120sec to 0.40 secs per page.
  
using the Benchmark class and running some test I got

normal: 0 0
    27.688000 0.015000 27.703000 ( 27.765000 )
fast:
    0.469000 0.000000 0.484000 (0.954000)

the speed difference is totally diferent.

Is this expected when using regular expressions??

regards,
Horacio

2

Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")

vs

   r1 = Regexp.new("#{key1}\.*#{key2}")
   r2 = Regexp.new("#{key2}\.*#{key1}")
   
Is this expected when using regular expressions??

yes, ruby has some optimizations. For example with the regexp /abc.*def/

svg% ruby -rjj -e '/abc.*def/.dump'
Regexp /abc.*def/
  0 exactn "abc" (3)
  1 anychar_repeat
  2 exactn "def" (3)
  3 end
must : abc
optimize : exactn
svg%

It call the regexp engine (which is slow) only when it has found the
substring "abc" in the string

Now if you use /abc.*def|def.*abc/ you break this optimization

svg% ruby -rjj -e '/abc.*def|def.*abc/.dump'
Regexp /abc.*def|def.*abc/
  0 on_failure_jump ==> 5
  1 exactn "abc" (3)
  2 anychar_repeat
  3 exactn "def" (3)
  4 jump ==> 8
  5 exactn "def" (3)
  6 anychar_repeat
  7 exactn "abc" (3)
  8 end
svg%

it must call the stupid (:-)) regexp engine for each line

Guy Decoux

3

Horacio Sanson wrote:

I have this little script that takes a list of keyword sets, each set
has only two keywords and for each one of them the script creates a
regular expression like this:

Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")

then I match it to a string that contains a long text fetched from a
web page.

a more complete pseudo-code

#########################################
long_text = get_web_page(url)

keyword_hash = load_keyword_array_from_database

keyword_hash.each_pair { |id, value|

    key1 = value[0]
    key2 = value[1]

    r = Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")
    return id if long_text =~ r
}

return -1
###########################################

Now this code works perfect, the problem is that the keyword_hash has
more than 300 elements and running this code can take between 50 to
120 seconds. Since I am processing more than 1000 pages with this
code it takes forever.

I solved this problem by replacing the regular expression match to

   r1 = Regexp.new("#{key1}\.*#{key2}")
   r2 = Regexp.new("#{key2}\.*#{key1}")

   return id if long_text =~ r1 or long_text =~ r2

I simply put the or statement outside the regular expresion and the
speedup was from 50~120sec to 0.40 secs per page.

using the Benchmark class and running some test I got

normal: 0 0
    27.688000 0.015000 27.703000 ( 27.765000
)
fast:
    0.469000 0.000000 0.484000 (0.954000)

the speed difference is totally diferent.

Is this expected when using regular expressions??

On obvious optimization is to create all regexps during
load_keyword_array_from_database() and not during iteration of the hash.
That way you just have to do it once and can reuse those regexps with
multiple pages you check.

Another possible optimization is to take your approach of splitting the
regexps a bit further and create two regexps - one for each keyword - and
return the id if both match. This works only correctly if (i) keywords
don't overlap or (ii) you can use \b to ensure matching on word
boundaries.

Kind regards

    robert

4

There's a whole section in Mastering Regular Expressions that goes into the
differences between regexp engines.

Summary: it makes a big difference how you setup your patterns!

···

On 11/22/05, Robert Klemme <bob.news@gmx.net> wrote:

Horacio Sanson wrote:
> I have this little script that takes a list of keyword sets, each set
> has only two keywords and for each one of them the script creates a
> regular expression like this:
>
> Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")
>
> then I match it to a string that contains a long text fetched from a
> web page.
>
> a more complete pseudo-code
>
> #########################################
> long_text = get_web_page(url)
>
> keyword_hash = load_keyword_array_from_database
>
> keyword_hash.each_pair { |id, value|
>
> key1 = value[0]
> key2 = value[1]
>
> r = Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")
> return id if long_text =~ r
> }
>
> return -1
> ###########################################
>
>
> Now this code works perfect, the problem is that the keyword_hash has
> more than 300 elements and running this code can take between 50 to
> 120 seconds. Since I am processing more than 1000 pages with this
> code it takes forever.
>
>
> I solved this problem by replacing the regular expression match to
>
> r1 = Regexp.new("#{key1}\.*#{key2}")
> r2 = Regexp.new("#{key2}\.*#{key1}")
>
> return id if long_text =~ r1 or long_text =~ r2
>
>
> I simply put the or statement outside the regular expresion and the
> speedup was from 50~120sec to 0.40 secs per page.
>
>
> using the Benchmark class and running some test I got
>
> normal: 0 0
> 27.688000 0.015000 27.703000 ( 27.765000
> )
> fast:
> 0.469000 0.000000 0.484000 (0.954000)
>
>
> the speed difference is totally diferent.
>
> Is this expected when using regular expressions??

On obvious optimization is to create all regexps during
load_keyword_array_from_database() and not during iteration of the hash.
That way you just have to do it once and can reuse those regexps with
multiple pages you check.

Another possible optimization is to take your approach of splitting the
regexps a bit further and create two regexps - one for each keyword - and
return the id if both match. This works only correctly if (i) keywords
don't overlap or (ii) you can use \b to ensure matching on word
boundaries.

Kind regards

robert

--
-

'There was an owl lived in an oak.
The more he heard, the less he spoke.
The less he spoke, the more he heard.'

Christian Leskowsky
christian.leskowsky@gmail.com

5

ts <decoux@moulon.inra.fr> writes:

> Regexp.new("#{key1}\.*#{key2}|#{key2}\.*#{key1}")

vs

> r1 = Regexp.new("#{key1}\.*#{key2}")
> r2 = Regexp.new("#{key2}\.*#{key1}")
   
> Is this expected when using regular expressions??

yes, ruby has some optimizations. For example with the regexp /abc.*def/

Do you know how oniguruma does that, per chance?

···

svg% ruby -rjj -e '/abc.*def/.dump'
Regexp /abc.*def/
  0 exactn "abc" (3)
  1 anychar_repeat
  2 exactn "def" (3)
  3 end
must : abc
optimize : exactn
svg%

--
Christian Neukirchen <chneukirchen@gmail.com> http://chneukirchen.org

6

Do you know how oniguruma does that, per chance?

You can compile oniguruma with debugging options

moulon% ./ruby -e '/abc.*def/'
<list:814d418>
   <string:814d3e8>abc
   <qualifier:814d478>{0,-1}
      <anychar:814d448>
   <string:814d4d8>def
optimize: EXACT_BM
  anchor:
  sub anchor:

exact: [abc]: length: 3
code length: 11
[exact3:abc] [anychar*-peek-next:d] [exact3:def] [end]
moulon%

Guy Decoux