handyfloss

As outlined in some previous posts[1,2,3,4], I have been playing around with a piece of Python code to process some log files. The log files to process were actually host.gz files from some [[BOINC]] projects, and the data I want to extract from them is quite simple: the Windows, Linux and Mac shares in the number of computers contributing to them (and the [[BOINC Credit System|work they do]]). By logging this processed data myself, I can see the time evolution of this share, and hopefully show the slow but steady rise of GNU/Linux :^)

I figured out that the contribution to distributed computing projects could be a reasonable indicator of the Windows predominance status. There are many other indicators (for example the number of visits to a web site, e.g. this very one), and I don’t claim that this one is “better”. I just want to add it to the reference list for the reader.

There is a problem with “Windows vs. Linux” figures, and it is that they are not really “competing” products. When cars or soft drinks are the subject, one can figure out the [[market share]], looking at the number of items sold. Linux being [[free software]], one can hardly measure the amount of “sold copies”, and with Windows being pre-installed in most new computers, one can not really trust the “number of computers sold = number of Windows copies sold”, because some users even remove the Windows partition and install Linux on top of it.

Counting the visits to some sites is not without problems, either. Any web site will have a particular audience, and the result will be biased by that fact. When my blog was in WordPress.com, I had roughly as many visits from Windows users as from Linux users, and almost all of them used Firefox as a browser. Obviously this data is not an accurate reflection of the world at large. It so happened that free software users are more likely to surf to sites like mine, hence the bias.

So, without further ado, let me introduce the “BOINC Host Statistics” program (BHS). Here you are a link to its home page. You can find results I have harvested so far in the Screenshots section. For example, the SETI@home credit generation rate statistics follows:

What the plot tells us is that (at the time of writing this) 500 million [[BOINC Credit System|cobblestones]] are being granted to contributors each day. Of them, around 82% are being given to Windows computers, 9-10% to Mac, 8% to GNU/Linux, and the rest to computers running other OSs.

Update: you can find the outcome of all this in a latter post: Project BHS

All the comments to my previous post have provided me with hints to increase further the efficiency of a script I am working on. Here I present the advices I have followed, and the speed gain they provided me. I will speak of “speedup”, instead of timing, because this second set of tests has been made in a different computer. The “base” speed will be the last value of my previous test set (1.5 sec in that computer, 1.66 in this one). A speedup of “2” will thus mean half an execution time (0.83 s in this computer).

Version 6: Andrew Dalke suggested the substitution of:

line = re.sub('>','<',line)

with:

line = line.replace('>','<')

Avoiding the re module seems to speed up things, if we are searching for fixed strings, so the additional features of the re module are not needed.

This is true, and I got a speedup of 1.37.

Version 7: Andrew Dalke also suggested substituting:

search_cre = re.compile(r'total_credit').search
if search_cre(line):

with:

if 'total_credit' in line:

This is more readable, more concise, and apparently faster. Doing it increases the speedup to 1.50.

Version 8: Andrew Dalke also proposed flattening some variables, and specifically avoiding dictionary search inside loops. I went further than his advice, even, and substituted:

stat['win'] = [0,0]

loop
  stat['win'][0] = something
  stat['win'][1] = somethingelse

with:

win_stat_0 = 0
win_stat_1 = 0

loop
  win_stat_0 = something
  win_stat_1 = somethingelse

This pushed the speedup futher up, to 1.54.

Version 9: Justin proposed reducing the number of times some patterns were matched, and extract some info more directly. I attained that by substituting:

loop:
  if 'total_credit' in line:
    line   = line.replace('>','<')
    aline  = line.split('<')
    credit = float(aline[2])

with:

pattern    = r'total_credit>([^<]+)<';
search_cre = re.compile(pattern).search

loop:
  if 'total_credit' in line:
    cre    = search_cre(line)
    credit = float(cre.group(1))

This trick saved enough to increase the speedup to 1.62.

Version 10: The next tweak was an idea of mine. I was diggesting a huge log file with zcat and grep, to produce a smaller intermediate file, which Python would process. The structure of this intermediate file is of alternating lines with “total_credit” then “os_name” then “total_credit”, and so on. When processing this file with Python, I was searching the line for “total_credit” to differentiate between these two lines, like this:

for line in f:
  if 'total_credit' in line:
    do something
  else:
    do somethingelse

But the alternating structure of my input would allow me to do:

odd = True
for line in f:
  if odd:
    do something
    odd = False
  else:
    do somethingelse
    odd = True

Presumably, checking falsity of a boolean is faster than matching a pattern, although in this case the gain was not huge: the speedup went up to 1.63.

Version 11: Another clever suggestion by Andrew Dalke was to avoid using the intermediate file, and use os.popen to connect to and read from the zcat/grep command directly. Thus, I substituted:

os.system('zcat host.gz | grep -F -e total_credit -e os_name > '+tmp)

f = open(tmp)
for line in f:
  do something

with:

f = os.popen('zcat host.gz | grep -F -e total_credit -e os_name')

for line in f:
  do something

This saves disk I/O time, and the performance is increased accordingly. The speedup goes up to 1.98.

All the values I have given are for a sample log (from MalariaControl.net) with 7 MB of gzipped info (49 MB uncompressed). I also tested my scripts with a 267 MB gzipped (1.8 GB uncompressed) log (from SETI@home), and a plot of speedups vs. versions follows:

Execution speedup vs. version
(click to enlarge)

Notice how the last modification (avoiding the temporary file) is of much more importance for the bigger file than for the smaller one. Recall also that the odd/even modification (version 10) is of very little importance for the small file, but quite efficient for the big file (compare it with Version 9).

The plot doesn’t tell (it compares versions with the same input, not one input with the other), but my eleventh version of the script runs the 267 MB log faster than the 7 MB one with Version 1! For the 7 MB input, the overall speedup from Version 1 to Version 11 is above 50.

This is a follow up to two previous posts. In the first one I spoke about saving memory by reading line-by-line, instead of all-at-once, and in the second one I recommended using Unix commands.

The script reads a host.gz log file from a given BOINC project (more precisely one I got from MalariaControl.net, because it is a small project, so its logs are also smaller), and extracts how many computers are running the project, and how much credit they are getting. The statistics are separated by operating system (Windows, Linux, MacOS and other).

Version 0

Here I read the whole file to RAM, then process it with Python alone. Running time: 34.1s.

#!/usr/bin/python

import os
import re
import gzip

credit  = 0
os_list = ['win','lin','dar','oth']

stat = {}
for osy in os_list:
  stat[osy] = [0,0]

# Process file:
f = gzip.open('host.gz','r')
for line in f.readlines():
  if re.search('total_credit',line):
    credit = float(re.sub('/?total_credit>',' ',line.split()[0])
  elif re.search('os_name',line):
    if re.search('Windows',line):
      stat['win'][0] += 1
      stat['win'][1] += credit
    elif re.search('Linux',line):
        stat['lin'][0] += 1
        stat['lin'][1] += credit
    elif re.search('Darwin',line):
      stat['dar'][0] += 1
      stat['dar'][1] += credit
    else:
      stat['oth'][0] += 1
      stat['oth'][1] += credit
f.close()

# Return output:
nstring = ''
cstring = ''
for osy in os_list:
  nstring +=   "%15.0f " % (stat[osy][0])
  try:
    cstring += "%15.0f " % (stat[osy][1])
  except:
    print osy,stat[osy]

print nstring
print cstring

Version 1

The only difference is a “for line in f:“, instead of “for line in f.readlines():“. This saves a LOT of memory, but is slower. Running time: 44.3s.

Version 2

In this version, I use precompiled regular expresions, and the time-saving is noticeable. Running time: 26.2s

#!/usr/bin/python

import os
import re
import gzip

credit  = 0
os_list = ['win','lin','dar','oth']

stat = {}
for osy in os_list:
  stat[osy] = [0,0]


pattern    = r'total_credit'
match_cre  = re.compile(pattern).match
pattern    = r'os_name';
match_os   = re.compile(pattern).match
pattern    = r'Windows';
search_win = re.compile(pattern).search
pattern    = r'Linux';
search_lin = re.compile(pattern).search
pattern    = r'Darwin';
search_dar = re.compile(pattern).search

# Process file:
f = gzip.open('host.gz','r')

for line in f:
  if match_cre(line,5):
    credit = float(re.sub('/?total_credit>',' ',line.split()[0])
  elif match_os(line,5):
    if search_win(line):
      stat['win'][0] += 1
      stat['win'][1] += credit
    elif search_lin(line):
      stat['lin'][0] += 1
      stat['lin'][1] += credit
    elif search_dar(line):
      stat['dar'][0] += 1
      stat['dar'][1] += credit
    else:
      stat['oth'][0] += 1
      stat['oth'][1] += credit
f.close()

# etc.

Version 3

Later I decided to use AWK to perform the heaviest part: parsing the big file, to produce a second, smaller, file that Python will read. Running time: 14.8s.

#!/usr/bin/python

import os
import re

credit  = 0
os_list = ['win','lin','dar','oth']

stat = {}
for osy in os_list:
  stat[osy] = [0,0]

pattern    = r'Windows';
search_win = re.compile(pattern).search
pattern    = r'Linux';
search_lin = re.compile(pattern).search
pattern    = r'Darwin';
search_dar = re.compile(pattern).search

# Distile file with AWK:
tmp = 'bhs.tmp'
os.system('zcat host.gz | awk \'/total_credit/{printf $0}/os_name/{print}\' > '+tmp)

stat = {}
for osy in os_list:
  stat[osy] = [0,0]
# Process tmp file:
f = open(tmp)
for line in f:
  line = re.sub('>','<',line)
  aline = line.split('<')
  credit = float(aline[2])
  os_str = aline[6]
  if search_win(os_str):
    stat['win'][0] += 1
    stat['win'][1] += credit
  elif search_lin(os_str):
    stat['lin'][0] += 1
    stat['lin'][1] += credit
  elif search_dar(os_str):
    stat['dar'][0] += 1
    stat['dar'][1] += credit
  else:
    stat['oth'][0] += 1
    stat['oth'][1] += credit
f.close()

# etc

Version 4

Instead of using AWK, I decided to use grep, with the idea that nothing can beat this tool, when it comes to pattern matching. I was not disappointed. Running time: 5.4s.

#!/usr/bin/python

import os
import re

credit  = 0
os_list = ['win','lin','dar','oth']

stat = {}
for osy in os_list:
  stat[osy] = [0,0]

pattern    = r'total_credit'
search_cre = re.compile(pattern).search

pattern    = r'Windows';
search_win = re.compile(pattern).search
pattern    = r'Linux';
search_lin = re.compile(pattern).search
pattern    = r'Darwin';
search_dar = re.compile(pattern).search

# Distile file with grep:
tmp = 'bhs.tmp'
os.system('zcat host.gz | grep -e total_credit -e os_name > '+tmp)

# Process tmp file:
f = open(tmp)
for line in f:
  if search_cre(line):
    line = re.sub('>','<',line)
    aline = line.split('<')
    credit = float(aline[2])
  else:
    if search_win(line):
      stat['win'][0] += 1
      stat['win'][1] += credit
    elif search_lin(line):
      stat['lin'][0] += 1
      stat['lin'][1] += credit
    elif search_dar(line):
      stat['dar'][0] += 1
      stat['dar'][1] += credit
    else:
      stat['oth'][0] += 1
      stat['oth'][1] += credit

f.close()

# etc

Version 5

I was not completely happy yet. I discovered the -F flag for grep (in the man page), and decided to use it. This flag tells grep that the pattern we are using is a literal, so no expansion of it has to be made. Using the -F flag I further reduced the running time to: 1.5s.

Running time vs. script version (Click to enlarge)

In my last post I commented some changes I made to a Python script to process a file reducing the memory overhead related to reading the file directly to RAM.

I realized that the script needed much optimizing, and resorted to reading the link a reader (Paddy3118) was kind enough to point me to, I realized I could save time by compiling my search expressions. Basically my script opens a gzipped file, searches for lines containing some keywords, and uses the info read from those lines. The original script would take 44 seconds to process a 6.9 MB file (49 MB uncompressed). Using compile on the search expressions, this time went down to 29 s. I tried using match instead of search, and expressions like “if pattern in line:“, instead of re.search(), but these didn’t make much of a difference.

Later I thought that Unix commands such as grep were specially suited for the task, so I gave them a try. I modified my script to run in two steps: in the first one I used zcat and awk (called from within the script) to create a much smaller temporary file with only the lines containing the information I wanted. In a second step, I would process this file with standard Python code. This hybrid approach reduced the processing time to just 12 s. Sometimes using the best tool really makes a difference, and it seems that the Unix utilities are hard to come close to in terms of performance.

It is only after programming exercises like this one that one realizes how important writing good code is (something I will probably never do, but I try). For some reason I always think of Windows, and how Microsoft refuses to make an efficient program, relying on improvementes on the hardware instead. It’s as if I tried to speed up my first script using a faster computer, instead of fixing the code to be more efficient.

I was making a script to process some log file, and I basically wanted to go line by line, and act upon each line if some condition was met. For the task of reading files, I generally use readlines(), so my first try was:

f = open(filename,'r')
for line in f.readlines():
  if condition:
    do something
f.close()

However, I realized that as the size of the file read increased, the memory footprint of my script increased too, to the point of almost halting my computer when the size of the file was comparable to the available RAM (1GB).

Of course, Python hackers will frown at me, and say that I was doing something stupid… Probably so. I decided to try a different thing to reduce the memory usage, and did the following:

f = open(filename,'r')
for line in f:
  if condition:
    do something
f.close()

Both pieces of code look very similar, but pay a bit of attention and you’ll see the difference.

The problem with “f.readlines()” is that it reads the whole file and assigns lines to the elements of an (anonymous, in this case) array. Then, the for loops through the array, which is in memory. This leads to faster execution, because the file is read once and then forgotten, but requires more memory, because an array of the size of the file has to be created in the RAM.

Fig. 1: Memory vs file size for both methods of reading the file

When you do “for line in f:“, you are effectively reading the lines one by one when you do each cycle of the loop. Hence, the memory use is effectively constant, and very low, albeit the disk is accessed more often, and this usually leads to slower execution of the code.

Fig. 2: Execution time vs file size for both methods of reading the file

I am a recent fan of Python, a very neat scripting language.

One thing I miss from Perl is the availability of labeled breaks. What are those? Suppose you have two nested loops. When a condition is met in the inner loop, you want to exit both loops. With Python there is not straightforward way of doing it. Imagine we are reading an array of data, line by line and column by column, and we want to exit when meeting the first zero value. With Perl:

LINELOOP: foreach my $i (0..$lines)
{
  COLLOOP: foreach my $j (0..$columns)
  {
     break LINELOOP unless $val[$i][$j];
  };
};

A simple “break” will exit the innermost loop, but we can use a label to exit a specific loop. However, in Python there is no such a thing as a labeled loop, as explained in this PEP.

My rant is with the explanations given by van Rossum himself in Python mailing list to reject the change:

1. The complexity added to the language, permanently.
2. My expectation that the feature will be abused more than it will be used right.

Wow! Incredible reasons!

The first one is silly: other languages have it, and it has worked fine. Adding complexity to a tool for the sake of it is really stupid, I agree. But the fact is labeled breaks would be tremendously useful, so the increase in complexity would be justified. Surely a language that can only print “Hello world” would be less complex, yet of little use.

The second reason is absolutely over-the-shoulder-of-the-users. So now good old Guido must guide his sheep along the “correct” path, lest we get lost! He is punishing the good programmers by not giving them a useful tool, so that bad programmers are protected from their stupidity. It’s like not selling cars at all because some people drive while drunk.

Just my 2 cents…

Siempre que he escuchado la canción Ciega, sordomuda de Shakira, me ha parecido que esta tía es una hacker y una friki.

[youtube=http://www.youtube.com/watch?v=KOeaxkMbGO8]

¿Por qué digo eso? Porque la primera estrofa de la canción…

Se me acaba el argumento
y la metodología
cada vez que se aparece
frente a mí tu anatomía

… habla claramente de un programa informático, por ejemplo un script Python:

#!/usr/bin/python

import sys

class myclass:

  def __init__(self,var):
   var = self.var

  #def mymethod(self):
  #  pass

if 'frente_a_mi' == 'tu_anatomia':
  sys.argv = []
  myclass.mymethod()