Hi there,
I have some scientific application written in python. There is a
good deal of list processing, but also some "simple" computation such
as basic linear algebra involved. I would like to speed things up
implementing some of the functions in C. So I need profiling.
I first tried to use the default python profiler, but profiling my
application multiplies the execution time by a factor between 10 and
100 ! So I decided to give a try to hotspot. I just followed the
example of the python library reference, but I have some strange
results concerning cpu time. My profiling script is something like the
following:
def run_foo():
print time.clock()
function_to_profile()
print time.clock()
prof = hotshot.Profile("essai.prof")
benchtime= prof.runcall(run_foo)
prof.close()
stats = hotshot.stats.load("essai.prof")
stats.strip_dirs()
stats.sort_stats('time', 'calls')
stats.print_stats(20)
The goal is to profile the function function_to_profile(). Running this
script gives me a CPU executime time of around 2 seconds, whereas the
difference between the two clock calls is around 10 seconds ! And I
don't run any other cpu consuming tasks at the same time, so this
cannot come from other running processes. Is there something perticular
about hotspot timing I should know ? I am not sure how I can get more
accurate results with hotspot.
I would appreciate any help,
Thanks 6 3725 co******@gmail.com writes: I have some scientific application written in python. There is a good deal of list processing, but also some "simple" computation such as basic linear algebra involved. I would like to speed things up implementing some of the functions in C. So I need profiling.
Why don't you use numarray for the linear algebra?
In message <11**********************@g47g2000cwa.googlegroups .com>, co******@gmail.com writes Hi there,
I have some scientific application written in python. There is a good deal of list processing, but also some "simple" computation such as basic linear algebra involved. I would like to speed things up implementing some of the functions in C. So I need profiling.
You haven't said which platform you are on. If you are on Windows you
may want to try Python Performance Validator. http://www.softwareverify.com
Stephen
--
Stephen Kellett
Object Media Limited http://www.objmedia.demon.co.uk/software.html
Computer Consultancy, Software Development
Windows C++, Java, Assembler, Performance Analysis, Troubleshooting co******@gmail.com wrote: Hi there,
I have some scientific application written in python. There is a good deal of list processing, but also some "simple" computation such as basic linear algebra involved. I would like to speed things up implementing some of the functions in C. So I need profiling.
I first tried to use the default python profiler, but profiling my application multiplies the execution time by a factor between 10 and 100 ! So I decided to give a try to hotspot.
OK - first of all, as someone else has asked, what platform are you
running? I'm assuming it's windows since you're referring to
time.clock() and then later saying "wall clock".
Next, what are you hoping that the profiler will give you? If you're
expecting it to give you the big picture of your application's
performance and give you "real runtime numbers", you may be
disappointed. It is a deterministic profiler and will give you CPU time
spent in different areas of code rather than and overall "how long did
this thing take to run?".
I just followed the example of the python library reference, but I have some strange results concerning cpu time. My profiling script is something like the following:
def run_foo(): print time.clock()
function_to_profile()
print time.clock()
prof = hotshot.Profile("essai.prof") benchtime= prof.runcall(run_foo) prof.close() stats = hotshot.stats.load("essai.prof") stats.strip_dirs() stats.sort_stats('time', 'calls') stats.print_stats(20)
Well, let's just add more confusion to the pot, shall we? Look at this
example (a slight hack from yours)::
import time
import hotshot
import hotshot.stats
def run_foo():
print time.clock()
print time.time()
time.sleep(5)
print time.clock()
print time.time()
prof = hotshot.Profile("essai.prof")
benchtime= prof.runcall(run_foo)
prof.close()
stats = hotshot.stats.load("essai.prof")
stats.strip_dirs()
stats.sort_stats('time', 'calls')
stats.print_stats(20)
and the output::
0.24
1126011669.55
0.24
1126011674.55
1 function calls in 0.000 CPU seconds
Ordered by: internal time, call count
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 0.000 0.000 tmphQNKbq.py:6(run_foo)
0 0.000 0.000 profile:0(profiler)
I inserted a time.time() call since I'm on Linux and time.clock()
returns a process's CPU time and wanted to show the "wall clock time" as
it were. So, the stats show 0 time taken, whereas time.time() shows 5
seconds. It's because the time.sleep() took a negligable amount of CPU
time which is what the profiler looks at.
The goal is to profile the function function_to_profile(). Running this script gives me a CPU executime time of around 2 seconds, whereas the difference between the two clock calls is around 10 seconds !
I would attribute the wall clock and profile time difference to the
overhead of hotshot. While hotshot is miles better than the "regular"
profiler, it can still take a little time to profile code.
And I don't run any other cpu consuming tasks at the same time, so this cannot come from other running processes. Is there something perticular about hotspot timing I should know ? I am not sure how I can get more accurate results with hotspot.
I would appreciate any help,
Thanks
HTH,
JMJ OK - first of all, as someone else has asked, what platform are you running? I'm assuming it's windows since you're referring to time.clock() and then later saying "wall clock".
Actually, no. I am working on a x86 linux (HT disabled for this
testing, as I thought it may introduce some subtilities). I am not sure
aht you mean by wall clock ? Next, what are you hoping that the profiler will give you? If you're expecting it to give you the big picture of your application's performance and give you "real runtime numbers", you may be disappointed. It is a deterministic profiler and will give you CPU time spent in different areas of code rather than and overall "how long did this thing take to run?".
I am not sure to understand the big difference between "time spent in
different areas of code" and "how long did this thing take to run?".
Looking at python doc for deterministic profiling, I understand the
implementation difference, and the performance implications, but I
don't see why deterministic profiling would not give me an overall
picture ?
Well, let's just add more confusion to the pot, shall we? Look at this example (a slight hack from yours)::
import time import hotshot import hotshot.stats
def run_foo(): print time.clock() print time.time()
time.sleep(5)
print time.clock() print time.time()
prof = hotshot.Profile("essai.prof") benchtime= prof.runcall(run_foo) prof.close() stats = hotshot.stats.load("essai.prof") stats.strip_dirs() stats.sort_stats('time', 'calls') stats.print_stats(20)
and the output::
0.24 1126011669.55 0.24 1126011674.55 1 function calls in 0.000 CPU seconds
Ordered by: internal time, call count
ncalls tottime percall cumtime percall filename:lineno(function) 1 0.000 0.000 0.000 0.000 tmphQNKbq.py:6(run_foo) 0 0.000 0.000 profile:0(profiler) I inserted a time.time() call since I'm on Linux and time.clock() returns a process's CPU time and wanted to show the "wall clock time" as it were. So, the stats show 0 time taken, whereas time.time() shows 5 seconds. It's because the time.sleep() took a negligable amount of CPU time which is what the profiler looks at.
Well, your example make actually more sense to me :) I understand the
difference between CPU time and time spent in the python code (even if
I was not clear in my previous post about it...). But this case does
not apply to my code, as my code is never "idled", takes 100 % of the
cpu, with no other CPU consuming task
I would attribute the wall clock and profile time difference to the overhead of hotshot. While hotshot is miles better than the "regular" profiler, it can still take a little time to profile code.
Well, if hotshot reported a timing which is longer than the execution
time without it, I would have considered that to be normal. Even in C,
using gprof has a non negligeable overhead, most of the time.
What I don't understand is why hotshot reports that do_foo is executed
in 2 seconds whereas it effectively takes more than 10 seconds ? Is it
because I don't understand what deterministic profiling is about ?
David co******@gmail.com wrote:
<snip> I am not sure to understand the big difference between "time spent in different areas of code" and "how long did this thing take to run?". Looking at python doc for deterministic profiling, I understand the implementation difference, and the performance implications, but I don't see why deterministic profiling would not give me an overall picture ?
I think from below you said you were more clear on this. Cool.
<snip>
Well, your example make actually more sense to me :) I understand the difference between CPU time and time spent in the python code (even if I was not clear in my previous post about it...). But this case does not apply to my code, as my code is never "idled", takes 100 % of the cpu, with no other CPU consuming task I would attribute the wall clock and profile time difference to the overhead of hotshot. While hotshot is miles better than the "regular" profiler, it can still take a little time to profile code.
Well, if hotshot reported a timing which is longer than the execution time without it, I would have considered that to be normal. Even in C, using gprof has a non negligeable overhead, most of the time.
Actually, I'd expect the opposite, but not as extreme for your case. I
would expect it to *report* that a piece of code took less time to
execute than I *observed* it taking. Reasons in the snipped area
above. Unless you're calling a C extension, in which case, IIRC, it's
supposed to report the actual execution time of the C call (and I guess
plus any overhead that hotshot may cause it to incur) in which case you
would be IMO 100% correct. I hope you're not calling a C extension, or
my head's gonna explode.
What I don't understand is why hotshot reports that do_foo is executed in 2 seconds whereas it effectively takes more than 10 seconds ? Is it because I don't understand what deterministic profiling is about ?
The profiler is supposed to be smart about how it tracks time spent in
execution so it doesn't get readings that are tainted by other processes
running or other stuff. I could easily see a 2->10 second disparity if
your process were idling somehow. Now, if you're doing a lot of IO, I
wonder if the profiler isn't taking into consideration any blocking
calls that may max out the CPU in IOWAIT... Are you doing a lot of IO?
David
JMJ
I was unhappy with both hotshot and the standard python profiler, so
I wrote my own, which may be what you are looking for. I've submitted
it as a patch at: http://sourceforge.net/tracker/index...70&atid=305470
It should add a minimum of overhead, give real numbers and also
gives stats on child calls. However, it is not compatible with
the stats module.
You can compile it as a standalone module. co******@gmail.com wrote: Hi there,
I have some scientific application written in python. There is a good deal of list processing, but also some "simple" computation such as basic linear algebra involved. I would like to speed things up implementing some of the functions in C. So I need profiling.
I first tried to use the default python profiler, but profiling my application multiplies the execution time by a factor between 10 and 100 ! So I decided to give a try to hotspot. I just followed the example of the python library reference, but I have some strange results concerning cpu time. My profiling script is something like the following:
def run_foo(): print time.clock()
function_to_profile()
print time.clock()
prof = hotshot.Profile("essai.prof") benchtime= prof.runcall(run_foo) prof.close() stats = hotshot.stats.load("essai.prof") stats.strip_dirs() stats.sort_stats('time', 'calls') stats.print_stats(20)
The goal is to profile the function function_to_profile(). Running this script gives me a CPU executime time of around 2 seconds, whereas the difference between the two clock calls is around 10 seconds ! And I don't run any other cpu consuming tasks at the same time, so this cannot come from other running processes. Is there something perticular about hotspot timing I should know ? I am not sure how I can get more accurate results with hotspot.
I would appreciate any help,
Thanks This thread has been closed and replies have been disabled. Please start a new discussion. Similar topics
by: mike420 |
last post by:
THE GOOD:
1. pickle
2. simplicity and uniformity
3. big library (bigger would be even better)
THE BAD:
|
by: Jon Perez |
last post by:
Can one run a 1.5 .pyc file with the 2.x version
interpreters and vice versa?
How about running a 2.x .pyc using a 2.y interpreter?
|
by: Maurice LING |
last post by:
This may be a dumb thing to ask, but besides the penalty for dynamic
typing, is there any other real reasons that Python is slower than Java?
maurice
|
by: dot |
last post by:
Hi all, I have written a Python huffman Encoding Module, for my own
amusement. I thought it might be educational/entertaining for other
people, so I've put it on my website and wrote about it a...
|
by: Ian Bicking |
last post by:
The Chicago Python User Group, ChiPy, will have its next meeting
on Thursday, December 9th, starting at 7pm. For more information on
ChiPy see http://chipy.org
Topic
-----
This month's...
|
by: 63q2o4i02 |
last post by:
Hi, I've been thinking about Python vs. Lisp. I've been learning
Python the past few months and like it very much. A few years ago I
had an AI class where we had to use Lisp, and I absolutely...
|
by: jwrweatherley |
last post by:
I'm pretty new to python, but am very happy with it. As well as using
it at work I've been using it to solve various puzzles on the Project
Euler site - http://projecteuler.net. So far it has not...
|
by: taylorcarr |
last post by:
A Canon printer is a smart device known for being advanced, efficient, and reliable. It is designed for home, office, and hybrid workspace use and can also be used for a variety of purposes. However,...
|
by: aa123db |
last post by:
Variable and constants
Use var or let for variables and const fror constants.
Var foo ='bar';
Let foo ='bar';const baz ='bar';
Functions
function $name$ ($parameters$) {
}
...
|
by: ryjfgjl |
last post by:
In our work, we often receive Excel tables with data in the same format. If we want to analyze these data, it can be difficult to analyze them because the data is spread across multiple Excel files...
|
by: BarryA |
last post by:
What are the essential steps and strategies outlined in the Data Structures and Algorithms (DSA) roadmap for aspiring data scientists? How can individuals effectively utilize this roadmap to progress...
|
by: nemocccc |
last post by:
hello, everyone, I want to develop a software for my android phone for daily needs, any suggestions?
|
by: Sonnysonu |
last post by:
This is the data of csv file
1 2 3
1 2 3
1 2 3
1 2 3
2 3
2 3
3
the lengths should be different i have to store the data by column-wise with in the specific length.
suppose the i have to...
|
by: marktang |
last post by:
ONU (Optical Network Unit) is one of the key components for providing high-speed Internet services. Its primary function is to act as an endpoint device located at the user's premises. However,...
|
by: Hystou |
last post by:
Most computers default to English, but sometimes we require a different language, especially when relocating. Forgot to request a specific language before your computer shipped? No problem! You can...
|
by: Oralloy |
last post by:
Hello folks,
I am unable to find appropriate documentation on the type promotion of bit-fields when using the generalised comparison operator "<=>".
The problem is that using the GNU compilers,...
| | |