[Numpy-discussion] Numpy Overhead

Discussion:

Sebastian K

2017-02-28 21:47:16 UTC

Hello everyone,

I'm interested in the numpy project and tried a lot with the numpy array.
I'm wondering what is actually done that there is so much overhead when I
call a function in Numpy. What is the reason?
Thanks in advance.

Regards

Sebastian Kaster

Benjamin Root

2017-02-28 22:03:18 UTC

Permalink

You are going to need to provide much more context than that. Overhead
compared to what? And where (io, cpu, etc.)? What are the size of your
arrays, and what sort of operations are you doing? Finally, how much
overhead are you seeing?

There can be all sorts of reasons for overhead, and some can easily be
mitigated, and others not so much.

Cheers!
Ben Root

Post by Sebastian K
Hello everyone,
I'm interested in the numpy project and tried a lot with the numpy array.
I'm wondering what is actually done that there is so much overhead when I
call a function in Numpy. What is the reason?
Thanks in advance.
Regards
Sebastian Kaster
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Sebastian K

2017-02-28 22:12:16 UTC

Permalink

Thank you for your answer.
For example a very simple algorithm is a matrix multiplication. I can see
that the heap peak is much higher for the numpy version in comparison to a
pure python 3 implementation.
The heap is measured with the libmemusage from libc:

*heap peak*
Maximum of all *size* arguments of malloc(3)
<http://man7.org/linux/man-pages/man3/malloc.3.html>, all products
of *nmemb***size* of calloc(3)
<http://man7.org/linux/man-pages/man3/calloc.3.html>, all *size*
arguments of
realloc(3)
<http://man7.org/linux/man-pages/man3/realloc.3.html>, *length*
arguments of mmap(2)
<http://man7.org/linux/man-pages/man2/mmap.2.html>, and *new_size*
arguments of mremap(2)
<http://man7.org/linux/man-pages/man2/mremap.2.html>.

Regards

Sebastian

Post by Benjamin Root
You are going to need to provide much more context than that. Overhead
compared to what? And where (io, cpu, etc.)? What are the size of your
arrays, and what sort of operations are you doing? Finally, how much
overhead are you seeing?
There can be all sorts of reasons for overhead, and some can easily be
mitigated, and others not so much.
Cheers!
Ben Root
On Tue, Feb 28, 2017 at 4:47 PM, Sebastian K <

_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Joseph Fox-Rabinovitz

2017-02-28 22:17:33 UTC

Permalink

It would really help to see the code you are using in both cases as well as
some heap usage numbers...

-Joe

Post by Sebastian K
Thank you for your answer.
For example a very simple algorithm is a matrix multiplication. I can see
that the heap peak is much higher for the numpy version in comparison to a
pure python 3 implementation.
*heap peak*
Maximum of all *size* arguments of malloc(3) <http://man7.org/linux/man-pages/man3/malloc.3.html>, all products
of *nmemb***size* of calloc(3) <http://man7.org/linux/man-pages/man3/calloc.3.html>, all *size* arguments of
realloc(3) <http://man7.org/linux/man-pages/man3/realloc.3.html>, *length* arguments of mmap(2) <http://man7.org/linux/man-pages/man2/mmap.2.html>, and *new_size*
arguments of mremap(2) <http://man7.org/linux/man-pages/man2/mremap.2.html>.
Regards
Sebastian

Post by Sebastian K
Hello everyone,
I'm interested in the numpy project and tried a lot with the numpy
array. I'm wondering what is actually done that there is so much overhead
when I call a function in Numpy. What is the reason?
Thanks in advance.
Regards
Sebastian Kaster
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Matthew Brett

2017-02-28 22:17:15 UTC

Permalink

Hi,

On Tue, Feb 28, 2017 at 2:12 PM, Sebastian K

Sebastian K

2017-02-28 22:57:49 UTC

Permalink

Yes it is true the execution time is much faster with the numpy function.

The Code for numpy version:

def createMatrix(n):
Matrix = np.empty(shape=(n,n), dtype='float64')
for x in range(n):
for y in range(n):
Matrix[x, y] = 0.1 + ((x*y)%1000)/1000.0
return Matrix

if __name__ == '__main__':
n = getDimension()
if n > 0:
A = createMatrix(n)
B = createMatrix(n)
C = np.empty(shape=(n,n), dtype='float64')
C = np.dot(A,B)

#print(C)

In the pure python version I am just implementing the multiplication with
three for-loops.

Measured data with libmemusage:
dimension of matrix: 100x100
heap peak pure python3: 1060565
heap peakt numpy function: 4917180

Post by Matthew Brett
Hi,
On Tue, Feb 28, 2017 at 2:12 PM, Sebastian K

Post by Sebastian K
Thank you for your answer.
For example a very simple algorithm is a matrix multiplication. I can see
that the heap peak is much higher for the numpy version in comparison to

Post by Sebastian K
pure python 3 implementation.
heap peak
Maximum of all size arguments of malloc(3), all

products

Post by Sebastian K
of nmemb*size of calloc(3), all size arguments of
realloc(3), length arguments of mmap(2), and new_size
arguments of mremap(2).

Could you post the exact code you're comparing?
I think you'll find that a naive Python 3 matrix multiplication method
is much, much slower than the same thing with Numpy, with arrays of
any reasonable size.
Cheers,
Matthew
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Joseph Fox-Rabinovitz

2017-02-28 23:00:07 UTC

Permalink

For one thing, `C = np.empty(shape=(n,n), dtype='float64')` allocates 10^4
extra elements before being immediately discarded.

-Joe

Post by Sebastian K
Yes it is true the execution time is much faster with the numpy function.
Matrix = np.empty(shape=(n,n), dtype='float64')
Matrix[x, y] = 0.1 + ((x*y)%1000)/1000.0
return Matrix
n = getDimension()
A = createMatrix(n)
B = createMatrix(n)
C = np.empty(shape=(n,n), dtype='float64')
C = np.dot(A,B)
#print(C)
In the pure python version I am just implementing the multiplication with
three for-loops.
dimension of matrix: 100x100
heap peak pure python3: 1060565
heap peakt numpy function: 4917180

Post by Matthew Brett
Hi,
On Tue, Feb 28, 2017 at 2:12 PM, Sebastian K

Post by Sebastian K
Thank you for your answer.
For example a very simple algorithm is a matrix multiplication. I can

see

Post by Sebastian K
that the heap peak is much higher for the numpy version in comparison

to a

Post by Sebastian K
pure python 3 implementation.
heap peak
Maximum of all size arguments of malloc(3), all

products

Post by Sebastian K
of nmemb*size of calloc(3), all size arguments of
realloc(3), length arguments of mmap(2), and new_size
arguments of mremap(2).

_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Sebastian K

2017-02-28 23:04:19 UTC

Permalink

Yes you are right. There is no need to add that line. I deleted it. But the
measured heap peak is still the same.

Post by Joseph Fox-Rabinovitz
For one thing, `C = np.empty(shape=(n,n), dtype='float64')` allocates 10^4
extra elements before being immediately discarded.
-Joe

Post by Matthew Brett
Hi,
On Tue, Feb 28, 2017 at 2:12 PM, Sebastian K

Post by Sebastian K
Thank you for your answer.
For example a very simple algorithm is a matrix multiplication. I can

see

Post by Sebastian K
that the heap peak is much higher for the numpy version in comparison

to a

Post by Sebastian K
pure python 3 implementation.
heap peak
Maximum of all size arguments of malloc(3), all

products

Post by Sebastian K
of nmemb*size of calloc(3), all size arguments of
realloc(3), length arguments of mmap(2), and new_size
arguments of mremap(2).

_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Matthew Brett

2017-02-28 23:18:07 UTC

Permalink

Hi,

On Tue, Feb 28, 2017 at 3:04 PM, Sebastian K

Post by Sebastian K
Yes you are right. There is no need to add that line. I deleted it. But the
measured heap peak is still the same.

You're applying the naive matrix multiplication algorithm, which is
ideal for minimizing memory use during the computation, but terrible
for speed-related stuff like keeping values in the CPU cache:

https://en.wikipedia.org/wiki/Matrix_multiplication_algorithm

The Numpy version is likely calling into a highly optimized compiled
routine for matrix multiplication, which can load chunks of the
matrices at a time, to speed up computation. If you really need
minimum memory heap usage and don't care about the order of
magnitude(s) slowdown, then you might need to use the naive method,
maybe implemented in Cython / C.

Cheers,

Matthew

Sebastian K

2017-02-28 23:24:53 UTC

Permalink

Thank you! That is the information I needed.

Post by Matthew Brett
Hi,
On Tue, Feb 28, 2017 at 3:04 PM, Sebastian K

Post by Sebastian K
Yes you are right. There is no need to add that line. I deleted it. But

the

Post by Sebastian K
measured heap peak is still the same.

You're applying the naive matrix multiplication algorithm, which is
ideal for minimizing memory use during the computation, but terrible
https://en.wikipedia.org/wiki/Matrix_multiplication_algorithm
The Numpy version is likely calling into a highly optimized compiled
routine for matrix multiplication, which can load chunks of the
matrices at a time, to speed up computation. If you really need
minimum memory heap usage and don't care about the order of
magnitude(s) slowdown, then you might need to use the naive method,
maybe implemented in Cython / C.
Cheers,
Matthew
_______________________________________________
NumPy-Discussion mailing list
https://mail.scipy.org/mailman/listinfo/numpy-discussion

Nathaniel Smith

2017-03-01 04:15:20 UTC

Permalink

On Feb 28, 2017 2:57 PM, "Sebastian K" <***@googlemail.com>
wrote:

Yes it is true the execution time is much faster with the numpy function.

The Code for numpy version:

def createMatrix(n):
Matrix = np.empty(shape=(n,n), dtype='float64')
for x in range(n):
for y in range(n):
Matrix[x, y] = 0.1 + ((x*y)%1000)/1000.0
return Matrix

if __name__ == '__main__':
n = getDimension()
if n > 0:
A = createMatrix(n)
B = createMatrix(n)
C = np.empty(shape=(n,n), dtype='float64')
C = np.dot(A,B)

#print(C)

In the pure python version I am just implementing the multiplication with
three for-loops.

Measured data with libmemusage:
dimension of matrix: 100x100
heap peak pure python3: 1060565
heap peakt numpy function: 4917180

4 megabytes is less than the memory needed just to load numpy :-). Try a
1000x1000 array (or even bigger), and I think you'll see more reasonable
results.

-n