By using this site, you agree to our updated Privacy Policy and our Terms of Use. Manage your Cookies Settings.
443,918 Members | 1,852 Online
Bytes IT Community
+ Ask a Question
Need help? Post your question and get tips & solutions from a community of 443,918 IT Pros & Developers. It's quick & easy.

Unexpected timing results with file I/O

P: n/a
After reading an earlier thread about opening and closing lots of files,
I thought I'd do a little experiment.

Suppose you have a whole lot of files, and you need to open each one,
append a string, then close them. There's two obvious ways to do it:
group your code by file, or group your code by procedure.

# Method one: grouped by file.
for each file:
open the file, append the string, then close it
# Method two: grouped by procedure.
for each file:
open the file
for each open file:
append the string
for each open file:
close the file
If you have N files, both methods make the same number of I/O calls: N
opens, N writes, N closes. Which is faster?

Intuitively, the first method has *got* to be faster, right? It's got one
loop instead of three and it doesn't build an intermediate list of open
file objects. It's so *obviously* going to be faster that it is hardly
worth bothering to check it with timeit, right?

Well, I wouldn't be writing unless that intuitive result was wrong. So
here's my test results:
Method 1:
>>import timeit
names = ['afile' + str(n) for n in range(1000)]
T = timeit.Timer('''for name in names:
.... fp = open(name, 'a'); fp.write('xyz\\n'); fp.close()
.... ''', 'from __main__ import names')
>>min(T.repeat(6, 500))
17.391216039657593
Method 2:
>>for name in names: # reset the files to an empty state.
.... fp = open(name, 'w'); fp.close()
....
>>T = timeit.Timer('''files = [open(name, 'a') for name in names]
.... for fp in files:
.... fp.write('xyz\\n')
.... for fp in files:
.... fp.close()
.... ''', '''from __main__ import names''')
>>min(T.repeat(6, 500))
16.823362112045288
Surprisingly, Method 2 is a smidgen faster, by about half a second over
500,000 open-write-close cycles. It's not much faster, but it's
consistent, over many tests, changing many of the parameters (e.g. the
number of files, the number of runs per timeit test, etc.).

I'm using Linux and Python 2.5.

So, what's going on? Can anyone explain why the code which does more work
takes less time?

--
Steven
Feb 4 '08 #1
Share this Question
Share on Google+
5 Replies


P: n/a
Steven D'Aprano wrote:
So, what's going on? Can anyone explain why the code which does more work
takes less time?
Short answer: CPU and RAM are much faster than hard disks.

The three loops and the creation of a list costs only a few CPU cycles
compared to flushing the new data to disk.

Christian

Feb 4 '08 #2

P: n/a
On Mon, 04 Feb 2008 15:17:18 +0000, Steven D'Aprano wrote:
# Method one: grouped by file.
for each file:
open the file, append the string, then close it
# Method two: grouped by procedure.
for each file:
open the file
for each open file:
append the string
for each open file:
close the file

Method 1:

17.391216039657593

Method 2:

16.823362112045288
Surprisingly, Method 2 is a smidgen faster, by about half a second over
500,000 open-write-close cycles. It's not much faster, but it's
consistent, over many tests, changing many of the parameters (e.g. the
number of files, the number of runs per timeit test, etc.).

I'm using Linux and Python 2.5.

So, what's going on? Can anyone explain why the code which does more work
takes less time?
Can't confirm this (Linux, Python 2.5):

Method 1: 15.380897998809814
Method 2: 18.085366010665894

I guess it's really all about the disk IO as my system monitor applet
shows that almost all of the time is spend in the kernel and very little
in user space.

Ciao,
Marc 'BlackJack' Rintsch
Feb 4 '08 #3

P: n/a
On Feb 4, 10:17 am, Steven D'Aprano <st...@REMOVE-THIS-
cybersource.com.auwrote:
After reading an earlier thread about opening and closing lots of files,
I thought I'd do a little experiment.

Suppose you have a whole lot of files, and you need to open each one,
append a string, then close them. There's two obvious ways to do it:
group your code by file, or group your code by procedure.

# Method one: grouped by file.
for each file:
open the file, append the string, then close it

# Method two: grouped by procedure.
for each file:
open the file
for each open file:
append the string
for each open file:
close the file

If you have N files, both methods make the same number of I/O calls: N
opens, N writes, N closes. Which is faster?

Intuitively, the first method has *got* to be faster, right? It's got one
loop instead of three and it doesn't build an intermediate list of open
file objects. It's so *obviously* going to be faster that it is hardly
worth bothering to check it with timeit, right?

Well, I wouldn't be writing unless that intuitive result was wrong. So
here's my test results:

Method 1:
>import timeit
names = ['afile' + str(n) for n in range(1000)]
T = timeit.Timer('''for name in names:

... fp = open(name, 'a'); fp.write('xyz\\n'); fp.close()
... ''', 'from __main__ import names')>>min(T.repeat(6, 500))

17.391216039657593

Method 2:
>for name in names: # reset the files to an empty state.

... fp = open(name, 'w'); fp.close()
...>>T = timeit.Timer('''files = [open(name, 'a') for name in names]

... for fp in files:
... fp.write('xyz\\n')
... for fp in files:
... fp.close()
... ''', '''from __main__ import names''')>>min(T.repeat(6, 500))

16.823362112045288

Surprisingly, Method 2 is a smidgen faster, by about half a second over
500,000 open-write-close cycles. It's not much faster, but it's
consistent, over many tests, changing many of the parameters (e.g. the
number of files, the number of runs per timeit test, etc.).

I'm using Linux and Python 2.5.

So, what's going on? Can anyone explain why the code which does more work
takes less time?

--
Steven


The code that does more work takes more time. The second one does
quite a bit less work. Think of it like this:

You have 500,000 people to fit through a door. Here are your options:

1. For each person, open the door, walk through the door, then close
the door.
2. Open the door, allow everyone to walk through, then close the
door.

Which one would you say would be a more efficient way to fit 500,000
people through the door?
Feb 4 '08 #4

P: n/a
En Mon, 04 Feb 2008 15:53:11 -0200, rdahlstrom <ro*************@gmail.com>
escribi�:
On Feb 4, 10:17 am, Steven D'Aprano <st...@REMOVE-THIS-
cybersource.com.auwrote:
>>
Suppose you have a whole lot of files, and you need to open each one,
append a string, then close them. There's two obvious ways to do it:
group your code by file, or group your code by procedure.

# Method one: grouped by file.
for each file:
open the file, append the string, then close it

# Method two: grouped by procedure.
for each file:
open the file
for each open file:
append the string
for each open file:
close the file

If you have N files, both methods make the same number of I/O calls: N
opens, N writes, N closes. Which is faster?
The code that does more work takes more time. The second one does
quite a bit less work. Think of it like this:

You have 500,000 people to fit through a door. Here are your options:

1. For each person, open the door, walk through the door, then close
the door.
2. Open the door, allow everyone to walk through, then close the
door.

Which one would you say would be a more efficient way to fit 500,000
people through the door?
Mmmm, no, the second one should be:

2. Create 500,000 doors and open them.
Make each person enter the room -one at a time- using its own door.
Close each of the 500,000 doors.

--
Gabriel Genellina

Feb 4 '08 #5

P: n/a
On Mon, 04 Feb 2008 17:08:02 +0000, Marc 'BlackJack' Rintsch wrote:
>Surprisingly, Method 2 is a smidgen faster, by about half a second over
500,000 open-write-close cycles. It's not much faster, but it's
consistent, over many tests, changing many of the parameters (e.g. the
number of files, the number of runs per timeit test, etc.).

I'm using Linux and Python 2.5.

So, what's going on? Can anyone explain why the code which does more
work takes less time?

Can't confirm this (Linux, Python 2.5):

Method 1: 15.380897998809814
Method 2: 18.085366010665894
Hmmm... does your system use software RAID? Mine does. I wonder if that's
a relevant factor?
I guess it's really all about the disk IO as my system monitor applet
shows that almost all of the time is spend in the kernel and very little
in user space.
I wouldn't be surprised if it was something to do with the OS caching
writes to disk. And saying that is really just me doing a lot of hand-
waving and saying "it's magic of what we know naught".

--
Steven
Feb 4 '08 #6

This discussion thread is closed

Replies have been disabled for this discussion.