Match beginning of two strings

[Ravi]

Hi,

I have about 200GB of data that I need to go through and extract the
common first part of a line. Something like this.
"abcdefghijklmnop"

Here I want to extract the common string "abcdefghijklmnop". Basically I
need a fast way to do that for any two given strings. For my situation,
the common string will always be at the beginning of both strings. I can
use regular expressions to do this, but from what I understand there is
a lot of overhead. New data is being generated at the rate of about 1GB
per hour, so this needs to be reasonably fast while leaving CPU time for
other processes.

Thanks
Ravi

 

[Jim Richardson]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi,

I have about 200GB of data that I need to go through and extract the
common first part of a line. Something like this.

"abcdefghijklmnop"

Here I want to extract the common string "abcdefghijklmnop". Basically I
need a fast way to do that for any two given strings. For my situation,
the common string will always be at the beginning of both strings. I can
use regular expressions to do this, but from what I understand there is
a lot of overhead. New data is being generated at the rate of about 1GB
per hour, so this needs to be reasonably fast while leaving CPU time for
other processes.

Thanks
Ravi

Are you trying to match any to any strings? or only a pair as above?


-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.2 (GNU/Linux)

iD8DBQE/LENWd90bcYOAWPYRAtWhAJ4ozTD1G3xLzVkeuJvPDJTsLbkcBQCfX4E0
YR/+zWSPDwX0uUf8y0QkxJs=
=sGTb
-----END PGP SIGNATURE-----

 

[Ravi]

While you can be forgiven for not have guessed, os.path is the place to
look:
import os.path
a = "abcdefghijklmnopqrstuvwxyz"
b = "abcdefghijklmnopBHLHT"
print os.path.commonprefix([a,b])

-Scott David Daniels
(e-mail address removed)

Certainly not where I was expecting it, Thanks

Ravi

 

[Ravi]

Are you trying to match any to any strings? or only a pair as above?

Just a pair at a time, and I only want the first N characters that are
common to both strings. The os.path.commonprefix works nicely. Thanks
for your help.

Ravi

 

[John Machin]

Hi,

I have about 200GB of data that I need to go through and extract the
common first part of a line. Something like this.

"abcdefghijklmnop"

Here I want to extract the common string "abcdefghijklmnop". Basically I
need a fast way to do that for any two given strings. For my situation,
the common string will always be at the beginning of both strings. I can
use regular expressions to do this, but from what I understand there is
a lot of overhead. New data is being generated at the rate of about 1GB
per hour, so this needs to be reasonably fast while leaving CPU time for
other processes.

Thanks
Ravi

While you can be forgiven for not have guessed, os.path is the place to
look:
import os.path
a = "abcdefghijklmnopqrstuvwxyz"
b = "abcdefghijklmnopBHLHT"
print os.path.commonprefix([a,b])

I don't think that os.path.commonprefix was designed with 200Gb of
data in mind. Inspection of Lib/*path.py gives one the impression that
it spends a lot of time discovering that the first element is a prefix
of itself.

Ravi, You may need to drop down to C to get the speed you want for
your requirement to find the longest common prefix of two strings. Two
things puzzling me: (1) how you would do this with regular expressions
(2) you have 200Gb now, new data arriving at the rate of 1Gb per hour,
after a year you have almost 9000Gb; where are you putting it all?
BTW, I do hope your algorithm is not O(N**2) ...

Cheers,
John

 

[Jeff Epler]

This is a naive implementation of the 'extract' function.
def extract(a, b):
m = min(len(a), len(b))
for i in range(m):
if a != b:
return a[:i]
return a[:m]

Here's one that uses the new zip() function:
def extract2(a, b):
m = min(len(a), len(b))
for i, ai, bi in (range(m), a, b):
if ai != bi: return a[:i]
return a[:m]
... unfortunately, it seems to be slower than the first method. On my
machine (800MHz PIII):
$ python timeit.py -s 'import ravi' \
'ravi.extract("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
10000 loops, best of 3: 32.7 usec per loop

If your goal is actually something slightly different---for instance,
find the string from a list with the largest shared prefix to a given
string---then you probably need to research an efficient algorithm.*
Otherwise, you may be stuck with the above. If you have 200GB, and each
line is 80 chars, you have about 2.7 billion lines. If you call extract()
once per line, and it takes 32 usec, you're looking at 24 hours to run.

Writing extract as a C extension may be wise, you're likely to be able to
cut those 32 usec down to little more than Python function call overhead,
or <2usec per loop. That makes your 2.7x10^9 calls only take 70 minutes.

Jeff
* The efficient algorithm for this problem might involve arranging the
list of strings as a tree, which makes finding the right prefix for
a given string against the list take only about lg(len(l)) steps,
instead of len(l) steps. This isn't so much a Python problem as a
computer science problem, though.

 

[Ravi]

I don't think that os.path.commonprefix was designed with 200Gb of

data in mind. Inspection of Lib/*path.py gives one the impression that
it spends a lot of time discovering that the first element is a prefix
of itself.

Ravi, You may need to drop down to C to get the speed you want for
your requirement to find the longest common prefix of two strings. Two
things puzzling me: (1) how you would do this with regular expressions
(2) you have 200Gb now, new data arriving at the rate of 1Gb per hour,
after a year you have almost 9000Gb; where are you putting it all?
BTW, I do hope your algorithm is not O(N**2) ...

Cheers,
John

Well, I timed os.path.commonprefix with some typical data and it's
pulling about 40usec per loop. So I did what I hated and coded a little
function in C. Goes something like this. My reasoning is that usually
the point where the strings start to differ is in the 30 - 50 character
range. Basically the idea is the same as a binary search on a sorted
array. Divide and conquer by going halfway each time.

Read in both strings.
Check to see if the first character matches.
If yes:
Check halfway through the string and see if that character matches
Repeatedly check halfway until the difference point is found.
Go back through from the difference point backwards and make sure
the characters match from the start to the difference point.

I timed it, and it seems to be doing about 3.5usec per loop. Using
pipes, I can feed it directly into the processing program. Good enough
for me.

As for the data, it's data from a radio telescope that's being recorded.
We do pattern analysis and reduce it to strings. By examining these
strings (more analysis than just the first common bit of course), we can
determine what data should be looked at further and what data is
garbage. The 9000GB problem isn't all that bad, the stuff compresses
extremely well, down to about 700GB for a year's worth. A couple of RAID
arrays makes quick work of that.

Thanks,

Ravi

 

[Jim Richardson]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Just a pair at a time, and I only want the first N characters that are
common to both strings. The os.path.commonprefix works nicely. Thanks
for your help.

Ravi

Well, it wasn't me, but you're welcome anyway




-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.2 (GNU/Linux)

iD8DBQE/LJrDd90bcYOAWPYRAlD9AJ9mqKMhx5if6eZ+i9vMuVkAsmfLugCgs2RE
A6o4gDz8qa/YhpjSoHdTwJo=
=ZAWM
-----END PGP SIGNATURE-----

 

[Andrew Dalke]

Jim Richardson:

Why bother finding out which one is the shorter? if you try the compare,
and you run out of the other to compare to, then by default, it's not
the same

Because I wasn't sure if the strings had embedded NULs in them.
Python strings allow those.

Otherwise something like this would work

char *s1, *s2 = ... the strings from Python
char *s = s1;
while ( *s1 && (*s1++ == *s2++))
;
return the string from s->s1, or just the size s1-s.

Andrew
(e-mail address removed)

 

[Bengt Richter]

This is a naive implementation of the 'extract' function.
def extract(a, b):
m = min(len(a), len(b))
for i in range(m):
if a != b:
return a[:i]
return a[:m]

Here's one that uses the new zip() function:

I don't see "zip" ;-)

def extract2(a, b):
m = min(len(a), len(b))
for i, ai, bi in (range(m), a, b):
if ai != bi: return a[:i]
return a[:m]

.. unfortunately, it seems to be slower than the first method. On my
machine (800MHz PIII):
$ python timeit.py -s 'import ravi' \
'ravi.extract("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
10000 loops, best of 3: 32.7 usec per loop

My timing harness (I seem to need a new getopt for timeit.py under 2.3)
shows a slight (15-22% less time) improvement for this 2.3 alternative:

def commonprefix(s1, s2): # very little tested!
try:
for i, c in enumerate(s1):
if c != s2: return s1[:i]
except IndexError:
return s1[:i]
return s1


[12:39] C:\pywk\clp>timefuns ravi -c extract -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmno
pBHLHT' -c commonprefix -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmnopBHLHT'
timing oh: 0.000007 ratio
extract: 0.000088 1.00
commonprefix: 0.000074 0.85

[12:39] C:\pywk\clp>timefuns ravi -c extract -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmno
pBHLHT' -c commonprefix -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmnopBHLHT'
timing oh: 0.000007 ratio
extract: 0.000091 1.00
commonprefix: 0.000071 0.78

[12:40] C:\pywk\clp>timefuns ravi -c extract -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmno
pBHLHT' -c commonprefix -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmnopBHLHT'
timing oh: 0.000007 ratio
extract: 0.000091 1.00
commonprefix: 0.000071 0.78

[12:40] C:\pywk\clp>timefuns ravi -c extract -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmno
pBHLHT' -c commonprefix -s 'abcdefghijklmnopqrstuvwxyz' -s 'abcdefghijklmnopBHLHT'
timing oh: 0.000007 ratio
extract: 0.000088 1.00
commonprefix: 0.000071 0.81

Regards,
Bengt Richter

 

[Jim Richardson]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Jim Richardson:

Because I wasn't sure if the strings had embedded NULs in them.
Python strings allow those.

Ah, I hadn't considered that.

Otherwise something like this would work

char *s1, *s2 = ... the strings from Python
char *s = s1;
while ( *s1 && (*s1++ == *s2++))
;
return the string from s->s1, or just the size s1-s.

Andrew
(e-mail address removed)

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.2 (GNU/Linux)

iD8DBQE/La14d90bcYOAWPYRApsqAJ9OW1jCViH/ytTMwpMv/TtTo9Q0BwCg2/Hn
R41XPXwJUhQYvCBnrqRMqLY=
=arqM
-----END PGP SIGNATURE-----

 

[Alex Martelli]

Jeff Epler wrote:
...

.. unfortunately, it seems to be slower than the first method. On my
machine (800MHz PIII):
$ python timeit.py -s 'import ravi' \
'ravi.extract("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
10000 loops, best of 3: 32.7 usec per loop

Here's my proposal...:

import sys

def extract(a, b):
m = min(len(a), len(b))
for i in range(m):
if a != b:
return a[:i]
return a[:m]

def extract2(a, b):
for i, ai, bi in zip(xrange(sys.maxint), a, b):
if ai != bi: return a[:i]
return a[:m]

def extract3(a, b):
for i, ai in enumerate(a):
if ai != b[i:i+1]:
return a[:i]
return a

[alex@lancelot python2.3]$ python -O timeit.py -s 'import exa'
'exa.extract("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
100000 loops, best of 3: 13.9 usec per loop
[alex@lancelot python2.3]$ python -O timeit.py -s 'import exa'
'exa.extract2("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
10000 loops, best of 3: 19.7 usec per loop
[alex@lancelot python2.3]$ python -O timeit.py -s 'import exa'
'exa.extract3("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
100000 loops, best of 3: 15.7 usec per loop

Now add after the "import sys" two lines:

import psyco
psyco.full()

and run again:

[alex@lancelot python2.3]$ python -O timeit.py -s 'import exa'
'exa.extract("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
1000000 loops, best of 3: 0.771 usec per loop

[alex@lancelot python2.3]$ python -O timeit.py -s 'import exa'
'exa.extract3("abcdefghijklmnopqrstuvwxyz","abcdefghijklmnopBHLHT")'
100000 loops, best of 3: 3.37 usec per loop

However, extract2 doesn't run correctly with psyco (gets a MemoryError).

Still, the 18-times acceleration that psyco is able to effect on
the naive extract IS typical of psyco's effect on functions coded
in simple, elementary terms. When you really need speed, assuming
that your processor is Intel-compatible, consider psyco (of course,
you'll generally use psyco.profile, or something more selective
still, rather than psyco.full) -- orders-of-magnitude improvements
on low-level bottleneck functions are anything but surprising...


Alex

 

[Alex Martelli]

Hi,

I have about 200GB of data that I need to go through and extract the
common first part of a line. Something like this.

"abcdefghijklmnop"

Here I want to extract the common string "abcdefghijklmnop". Basically I
need a fast way to do that for any two given strings. For my situation,
the common string will always be at the beginning of both strings. I can

Here's my latest study on this:

*** pexa.py:

import sys

import psyco
psyco.full()

import cexa
import exa

def extract(a, b):
m = min(len(a), len(b))
for i in range(m):
if a != b:
return a[:i]
return a[:m]

def extract2(a, b):
for i, ai, bi in zip(xrange(len(a)), a, b):
if ai != bi: return a[:i]
return a[:m]

def extract3(a, b):
for i, ai in enumerate(a):
if ai != b[i:i+1]:
return a[:i]
return a

extract_pyrex = exa.exa

extract_c = cexa.cexa

*** exa.pyx:

def exa(a, b):
cdef int la
cdef int lb
la = len(a)
lb = len(b)
cdef int lmin
lmin = min(la, lb)
cdef int i
i = 0
while i < lmin:
if a != b:
return a[:i]
i = i + 1
if lmin == la:
return a
else:
return b

*** cexa.c:

#include <Python.h>

static PyObject*
cexa(PyObject* self, PyObject* args)
{
char *a, *b;
int la, lb;
int lmin, i;
if(!PyArg_ParseTuple(args, "s#s#", &a, &la, &b, &lb))
return 0;

lmin = la;
if(lmin<lb) lmin = lb;

for(i=0; i<lmin; i++)
if(a != b)
break;

return Py_BuildValue("s#", a, i);
}

static PyMethodDef cexaMethods[] = {
{"cexa", cexa, METH_VARARGS, "Extract common prefix"},
{0}
};

void
initcexa(void)
{
Py_InitModule("cexa", cexaMethods);
}

I've built the pyrex-coded extension with:

from distutils.core import setup
from distutils.extension import Extension
from Pyrex.Distutils import build_ext

setup(
name = "exa",
ext_modules=[
Extension("exa", ["exa.pyx"])
],
cmdclass = {'build_ext': build_ext}
)

and the C-coded one with:

from distutils.core import setup
from distutils.extension import Extension

setup(
name = "cexa",
ext_modules=[
Extension("cexa", ["cexa.c"])
],
)

and my measurements give me:

[alex@lancelot exi]$ python -O timeit.py -s 'import pexa' \

'pexa.extract("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'

100000 loops, best of 3: 2.39 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
100000 loops, best of 3: 2.14 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract2("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
10000 loops, best of 3: 30.2 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract3("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
100000 loops, best of 3: 9.59 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract_pyrex("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
10000 loops, best of 3: 21.8 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract_c("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
100000 loops, best of 3: 1.88 usec per loop
[alex@lancelot exi]$

So, it seems you can still get a tiny drop of extra speed
with a C-coded extension, though it's doubtful whether it's
worth the bother wrt the pyro-optimized simple Python code
in function 'extract'. I'm not sure where I went wrong in
the Pyrex coding (it doesn't seem to be performing anywhere
as well as I thought it might) and I'll be happy for real
Pyrex expert to show me the way.

Of course, as others have pointed out, it's unclear from
your problem description that doing such operations pairwise
on a lot of pairs of strings is actually what you need. It
IS quite possible that what you're doing could often be
better modeled, e.g., by repeated "prefix extractions"
between ONE fixed string and several other candidate strings;
or "prefix extraction" between a set of more than two strings.
In each case, it's likely that you can get much better
performance by more sophisticated algorithms. However,
which algorithms those might be is unclear unless you can
provide mode details on what you're doing.


Alex

 

[John Machin]

I'm not sure where I went wrong in
the Pyrex coding (it doesn't seem to be performing anywhere
as well as I thought it might) and I'll be happy for real
Pyrex expert to show me the way.

I don't call myself an expert, but here's my best shot:

If you look at the generated C code, you'll see lots of conversion
between C and Python types. The trick is to get your args into C, stay
in C as much as possible, and ship back a Python return value.

It's made harder with strings as there is not (yet) any way of hinting
to Pyrex to use the "s#" gadget, you have to DIY, see below.

cdef extern from "Python.h":
int PyString_Size(object s)

def exa2(arga, argb):
cdef int la, lb, lmin, i
cdef char *a, *b

a = arga
b = argb
la = PyString_Size(arga)
lb = PyString_Size(argb)
# living dangerously, not testing for error;
# Easy to eyeball for correctness in this case,
# but ...
if la <= lb:
lmin = la
else:
lmin = lb
i = 0
while i < lmin:
if a != b:
return arga[:i]
i = i + 1
if lmin == la:
return arga
else:
return argb

 

[Richie Hindle]

[Alex]

I'm not sure where I went wrong in
the Pyrex coding (it doesn't seem to be performing anywhere
as well as I thought it might) and I'll be happy for real
Pyrex expert to show me the way.

I'm no an expert, but I can see a few easily-fixed problems. The line:

if a != b:

is working with Python strings when it could be working with C
strings. Here's the original code and its output on my machine:

def exa(a, b):
cdef int la
cdef int lb
la = len(a)
lb = len(b)
cdef int lmin
lmin = min(la, lb)
cdef int i
i = 0
while i < lmin:
if a != b:
return a[:i]
i = i + 1
if lmin == la:
return a
else:
return b

100000 loops, best of 3: 9.11 usec per loop

Here's a modified version of the code comparing C strings:

def exa(a, b):
cdef char* c_a # `a` as a C string
cdef char* c_b # `b` as a C string
cdef int la
cdef int lb

c_a = a
c_b = b
la = len(a)
lb = len(b)
cdef int lmin
lmin = min(la, lb)
cdef int i
i = 0
while i < lmin:
if c_a != c_b:
return a[:i]
i = i + 1
if lmin == la:
return a
else:
return b

100000 loops, best of 3: 5.79 usec per loop

Almost twice as fast. Looking at the generated C is always
worthwhile when optimising Pyrex code - here's the code that does
the comparison against Python strings:

/* "D:\src\tests\pyrex\exa.pyx":11 */
__pyx_3 = PyInt_FromLong(__pyx_v_i); if (!__pyx_3) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 11; goto __pyx_L1;}
__pyx_1 = PyObject_GetItem(__pyx_v_a, __pyx_3); if (!__pyx_1) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 11; goto __pyx_L1;}
Py_DECREF(__pyx_3); __pyx_3 = 0;
__pyx_5 = PyInt_FromLong(__pyx_v_i); if (!__pyx_5) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 11; goto __pyx_L1;}
__pyx_2 = PyObject_GetItem(__pyx_v_b, __pyx_5); if (!__pyx_2) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 11; goto __pyx_L1;}
Py_DECREF(__pyx_5); __pyx_5 = 0;
if (PyObject_Cmp(__pyx_1, __pyx_2, &__pyx_4) < 0) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 11; goto __pyx_L1;}
__pyx_4 = __pyx_4 != 0;
Py_DECREF(__pyx_1); __pyx_1 = 0;
Py_DECREF(__pyx_2); __pyx_2 = 0;
if (__pyx_4) {

vs.

/* "D:\src\tests\pyrex\exa.pyx":16 */
__pyx_5 = ((__pyx_v_c_a[__pyx_v_i]) != (__pyx_v_c_b[__pyx_v_i]));
if (__pyx_5) {

for C strings. There's another similar optimisation that the C
output leads you to: you can use strlen rather than Python's len:

cdef extern from "string.h":
int strlen(char*)

def exa(a, b):
cdef char* c_a # `a` as a C string
cdef char* c_b # `b` as a C string
cdef int la
cdef int lb

c_a = a
c_b = b
la = strlen(c_a)
lb = strlen(c_b)
cdef int lmin
lmin = min(la, lb)
cdef int i
i = 0
while i < lmin:
if c_a != c_b:
return a[:i]
i = i + 1
if lmin == la:
return a
else:
return b

100000 loops, best of 3: 3.58 usec per loop

That replaces:

/* "D:\src\tests\pyrex\exa.pyx":4 */
__pyx_1 = __Pyx_GetName(__pyx_b, "len"); if (!__pyx_1) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 4; goto __pyx_L1;}
__pyx_2 = PyTuple_New(1); if (!__pyx_2) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 4; goto __pyx_L1;}
Py_INCREF(__pyx_v_a);
PyTuple_SET_ITEM(__pyx_2, 0, __pyx_v_a);
__pyx_3 = PyObject_CallObject(__pyx_1, __pyx_2); if (!__pyx_3) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 4; goto __pyx_L1;}
Py_DECREF(__pyx_1); __pyx_1 = 0;
Py_DECREF(__pyx_2); __pyx_2 = 0;
__pyx_4 = PyInt_AsLong(__pyx_3); if (PyErr_Occurred()) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 4; goto __pyx_L1;}
Py_DECREF(__pyx_3); __pyx_3 = 0;
__pyx_v_la = __pyx_4;

/* "D:\src\tests\pyrex\exa.pyx":5 */
__pyx_1 = __Pyx_GetName(__pyx_b, "len"); if (!__pyx_1) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 5; goto __pyx_L1;}
__pyx_2 = PyTuple_New(1); if (!__pyx_2) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 5; goto __pyx_L1;}
Py_INCREF(__pyx_v_b);
PyTuple_SET_ITEM(__pyx_2, 0, __pyx_v_b);
__pyx_3 = PyObject_CallObject(__pyx_1, __pyx_2); if (!__pyx_3) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 5; goto __pyx_L1;}
Py_DECREF(__pyx_1); __pyx_1 = 0;
Py_DECREF(__pyx_2); __pyx_2 = 0;
__pyx_4 = PyInt_AsLong(__pyx_3); if (PyErr_Occurred()) {__pyx_filename = __pyx_f[0]; __pyx_lineno = 5; goto __pyx_L1;}
Py_DECREF(__pyx_3); __pyx_3 = 0;
__pyx_v_lb = __pyx_4;

with :

/* "D:\src\tests\pyrex\exa.pyx":12 */
__pyx_v_la = strlen(__pyx_v_c_a);

/* "D:\src\tests\pyrex\exa.pyx":13 */
__pyx_v_lb = strlen(__pyx_v_c_b);

and leaves the call to 'min' as the only remaining huge block of C.
The final version looks like this, eliminating 'min' (Greg, can we have
the terary operator in Pyrex please? <good_mood ? wink : frown>)

cdef extern from "string.h":
int strlen(char*)

def exa(a, b):
cdef char* c_a # `a` as a C string
cdef char* c_b # `b` as a C string
cdef int la
cdef int lb

c_a = a
c_b = b
la = strlen(c_a)
lb = strlen(c_b)
cdef int lmin
if la < lb:
lmin = la
else:
lmin = lb
cdef int i
i = 0
while i < lmin:
if c_a != c_b:
return a[:i]
i = i + 1
if lmin == la:
return a
else:
return b


1000000 loops, best of 3: 0.803 usec per loop

Over ten times quicker than the original, for the sake of a couple of
small tweaks driven by looking at the C output. Although the C still
looks very verbose at first glance, it's now substantially the same as
Alex's cexa.c.

Hope that helps,

 

[Bengt Richter]

Here's my latest study on this:

*** pexa.py:

[...]

JFTHOI, if you have the inclination, I'm curious how this slightly
different 2.3-dependent version would fare in your harness on your
system with the rest:

def commonprefix(s1, s2): # very little tested!
try:
for i, c in enumerate(s1):
if c != s2: return s1[:i]
except IndexError:
return s1[:i]
return s1

[...]

and my measurements give me:

[alex@lancelot exi]$ python -O timeit.py -s 'import pexa' \

'pexa.extract("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'

100000 loops, best of 3: 2.39 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
100000 loops, best of 3: 2.14 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract2("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
10000 loops, best of 3: 30.2 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract3("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
100000 loops, best of 3: 9.59 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract_pyrex("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
10000 loops, best of 3: 21.8 usec per loop
[alex@lancelot exi]$ python -O timeit.py -s 'import pexa'
'pexa.extract_c("abcdefghijklmonpKOU", "abcdefghijklmonpZE")'
100000 loops, best of 3: 1.88 usec per loop
[alex@lancelot exi]$

Interesting, but I think I will have to write a filter so I can
see a little more easily what your timeit.py outputs say ;-)

Regards,
Bengt Richter

 

[Ravi]

Hi,

I have about 200GB of data that I need to go through and extract the
common first part of a line. Something like this.

"abcdefghijklmnop"

Here I want to extract the common string "abcdefghijklmnop". Basically I
need a fast way to do that for any two given strings. For my situation,
the common string will always be at the beginning of both strings. I can
use regular expressions to do this, but from what I understand there is
a lot of overhead. New data is being generated at the rate of about 1GB
per hour, so this needs to be reasonably fast while leaving CPU time for
other processes.

Thanks
Ravi

I really appreciate all your help, Alex, Jim, Jeff, Andrew, John, Richie
and Bengt. However I have this problem taken care of now. Took around 6
hours to run on a P4 2.8Ghz 1.0GB DDR (I suspect I/O limitations). As
for the data, if you want to know about it just for the sake of an
optimized algorithm, there are no Null (\0) characters in the strings
(actually they're Base64), and I've included a typical pair of strings.
The version I used was Andrew's.

Someone suggested that this would be better done in larger sets than
just pairs. That's not suitable because of the structure of the data,
two strings might be highly correlated, but are probably quite different
from another pair of strings. Perhaps more significantly, correlation in
sets of greater than two has no physical significance to the experiment.

I grabbed this from a typical data file. So I would want to be
extracting 'A832nv81a'
"
A832nv81a81nW103v9c24jgpy92T
A832nv81aTyqiep4v9c324jgpy92T
"

Thanks for your help everyone, coming from a Perl (It's a four letter
word to me world, I'm very impressed by how helpful all of you are.

Ravi

 

[John Machin]

for C strings. There's another similar optimisation that the C
output leads you to: you can use strlen rather than Python's len:

You can, if you don't care about the possibility that the input may contain NULs.

 

[Bengt Richter]

I really appreciate all your help, Alex, Jim, Jeff, Andrew, John, Richie
and Bengt. However I have this problem taken care of now. Took around 6
hours to run on a P4 2.8Ghz 1.0GB DDR (I suspect I/O limitations). As
for the data, if you want to know about it just for the sake of an
optimized algorithm, there are no Null (\0) characters in the strings
(actually they're Base64), and I've included a typical pair of strings.
The version I used was Andrew's.

Someone suggested that this would be better done in larger sets than
just pairs. That's not suitable because of the structure of the data,
two strings might be highly correlated, but are probably quite different
from another pair of strings. Perhaps more significantly, correlation in
sets of greater than two has no physical significance to the experiment.

I grabbed this from a typical data file. So I would want to be
extracting 'A832nv81a'
"
A832nv81a81nW103v9c24jgpy92T
A832nv81aTyqiep4v9c324jgpy92T
"

I still don't understand your use case ;-)

1) Are you periodically batch processing to look for near-pairs in a 200GB
unsorted list of strings? Are they just in a huge unsorted ascii file with
line feeds delimiting?

or

1a) Does someone send you a single string in a message every now and then (how often?)
and ask you to find the closest match in the data base? I.e., 200GB at 30 bytes/string
would be 6.67 billion strings. Which you say you now crunch in ~6hrs? You don't do that
amount of crunching just for one match request, do you?

1b) If you get a lot of "match requests," wouldn't you gain a lot by at least partially
ordering the incoming streams into separate buckets? E.g., the simplest thing would be
to have 64 files corresponding to the first character, or even 64*64 files for the first two.
Then have 64*64 file buffers (not open files) of say 1000 strings each, which would be
64*64*1000*30 or call it 32k/file buffer -> 64*64*32*1024/(1024*1024) ->128 megabytes of buffers
and on the average you'd expect 1GB/hr to fill a buffer of 32k about
(1e9/3600)/(32*1024) -> 8.477 times/second. It should be reasonable to open a file for append
and write 32k and close it that often, IWT. Whateve box writes the 200GB data now could either
do the partitioning itself or send it by gigabit ethernet to a dedicated box for that, IWT. Or
you could distribute the load by patitioning the ethernet destinations per the leading n bits
and let 2**n boxes maybe do even more ordering on the fly.

Even without distributing the load, this could give you 200GB/4096 (if evenly distributed!!)
or only '%e'%(200e9/4096) ->4.882813e+007 or less than 50MB to read off disk and search
to make a probe with a single string. If you sorted when you probed and wrote back the
sorted data with an indication that it was sorted, further probes to that same partition could
go a lot faster.

2) Why do the string lengths vary if they are samples from some process? Are they
effectively just right-stripped from some fixed length, which would be a guaranteed max length?
2a) what are the max and min lengths?

3) what is the distribution of leading characters as they come from the source?
3a) E.g., is the first character uniformly distributed within its possible Base64 codes?
3b) Are they uncorrelated timewise? Or e.g. do they arrive in clumps of <<64 possibilities for a time?
3c) Are the individual strings' characters randomly distributed and uncorrelated within the string?
3d) You say 9000GB compresses to 700GB. That suggests a lot of correlation and/or uneven distributions.
Is there a fixed dictionary you could use to repack the strings bit-wise with huffman and/or
rle compression?
3d1) What is the significance of the Base64 character boundaries? I.e., would a common prefix of
8-bit bytes representing sequentially compressed string be good enough, even if the first non-
matching byte contained some bits that did match and would have been included in base64?

3d2) Note that compressing during partitioned 4096-bucket storage could save a lot of space
as well as speed up matching.

4) 1GB/hr translates to about 10k strings like the your example per second.
4a) Are they just logged sequentially to a 200GB data store? (cf. 1b)
4b==2a) Is there a max and/or min length to these strings? (the above are 29 & 30 chars).
4c) You say they are base 64. That means binary would make (6/8)*200gb = 150GB, and the
strings (6/8)*30 ~ 22.5 or say 24 for a nice multiple of 8, even without compression.

Just some thoughts. Might want to check my arithmetic ;-)

Thanks for your help everyone, coming from a Perl (It's a four letter
word to me world, I'm very impressed by how helpful all of you are.

Some newsgroups are like that. In the past I spent a fair amount of time on the Borland
Delphi n.g., and that was mostly very friendly and helpful too. I think that's just the
way people are unless some stupidities get in the way. Renaissance yes! Armageddon no!

Regards,
Bengt Richter

 

[=?ISO-8859-1?Q?Ma=F1ungo?=]

Ravi:

Read in both strings.
Check to see if the first character matches.
If yes:
Check halfway through the string and see if that character matches
Repeatedly check halfway until the difference point is found.
Go back through from the difference point backwards and make sure
the characters match from the start to the difference point.

I timed it, and it seems to be doing about 3.5usec per loop.

There's a lot of overhead for doing that. Have you tried the simple

char *s1 = ... the first string ..
char *s2 = ... the second string ..
n = ... the shorter of the two ..
for(i=0; i<n; i++) {
if (*s1++ != *s2++) {
break;
}
}
return ... the string s1[:n] (or even just the int)

Easy to understand, and the CPU is spending almost its whole
time doing character tests.

Andrew
(e-mail address removed)

First, I'm not a python programmer... but I think is better test int.
Something like that:

int *a = (int *) ...first string...;
int *b = (int *) ...second string...;

for( i = 0; i < n; i += 4 )
if( *a++ != *b++ )
break;
char *aa = (char *) a;
char *bb = (char *) b;
if( *aa++ == *bb++ ) i++;
if( *aa++ == *bb++ ) i++;
if( *aa == *bb ) i++;

and return i.