[Numpy-discussion] is numerical python only for prototyping ?
verveer at embl-heidelberg.de
Tue Jul 26 10:15:15 EDT 2005
On 26 Jul 2005, at 18:41, Sebastian Haase wrote:
> This is not sopposed to be an evil question; instead I'm hoping for the
> answer: "No, generally we get >=95% the speed of a pure C/fortran
> implementation" ;-)
you won't, generally. Question is: since you are certainly not going to
gain an order of a magnitude doing it in C, do you really care?
> But as I am the strongest Python/numarray advocate in our group I get
> the answer that Matlab is (of course) also very convenient but
> memory handling and overall execution performance is so bad that for
> implementation one would generally have to reimplement in C.
Well its true that implementations in C will be faster. And memory
handling in Numeric/numarray can be a pain since the tendency is to
create and destroy a lot of arrays if you are not careful.
> We are a bio-physics group at UCSF developping new algorithms for
> deconvolution (often in 3D). Our data sets are regularly bigger than
> 100MB. When deciding for numarray I was assuming that the "Hubble
> Crowd" had
> a similar situation and all the operations are therefore very much
> for this type of data.
Funny you mention that example. I did my PhD in exactly the same field
(considering you are from Sedats lab I guess you are in exactly the
same field as I was/am, i.e. fluorescence microscopy. What are you guys
up to these days?) and I developed all my algorithms in C at the time.
Now, about 7 years later, I returned to the field to re-implement and
extend some of my old algorithms for use with microscopy data that can
consist of multiple sets, each several 100MB at least. Now I use python
with numarray, and I am actualy quite happy with. I am pushing it by
using up to 2GB of memory, (per process, after splitting the problem
up and distributing it on a cluster...), but it works. I am sure I
could squeeze maybe a factor of two or three in terms of speed and
memory usage by rewriting in C, but that is currently not worth my
time. So I guess that counts as using numarray as a prototyping
environment, but the result is also suitable for production use.
> Is 95% a reasonable number to hope for ? I did wrap my own version of
> (with "plan-caching"), which should give 100% of the C-speed.
That should help a lot, as the standard FFTs that come with
Numarray/Numeric suck big time. I do use them, but have to go through
all kind of tricks to get some decent memory usage in 32bit floating
point. The FFT array module is in fact very badly written for use with
large multi-dimensional data sets.
> But concerns
> arise from expression like "a=b+c*a" (think "convenience"!): If a,b,c
> each 3D-datastacks creation of temporary data-arrays for 'c*a' AND
> then also
> for 'b+...' would have to be very costly. (I think this is at least
> for Numeric - I don't know about Matlab and numarray)
That is indeed a problem, although I think in your case you maybe
limited by your FFTs anyway, at least in terms of speed. One thing you
should consider is replacing expressions such as ' c= a + b' with
add(a, b, c). If you do that cleverly you can avoid quite some memory
allocations and you 'should' get closer to C. That does not solve
1) Complex expressions still need to be broken up in sequences of
operations which is likely slower then iterating once over you array
and do the expression at each point.
2) Unfortunately not all numarray functions support an output array
(maybe only the ufuncs?). This can be a big problem, as then temporary
arrays must be allocated. (It sure was a problem for me.)
You can of course always re-implement the parts that are critical in C
and wrap them (as you did with FFTW). In fact, I think numarray now
provides a relatively easy way to write ufuncs which would allow you to
write a single python function in C for complex expressions.
> Hoping for comments,
Hope this gives some insights. I guess I have had similar experiences,
there are definitely some limits to the use of numarray/Numeric that
could be relieved, for instance by having a consistent implementation
of output arrays. That would allow writing algorithms where you could
very strictly control the allocation and de-allocation of arrays, which
would be a big help for working with large arrays.
I would not mind to hear a bit about your experiences doing the big
deconvolutions in numarray/Numeric, but that may not be a good topic
for this list.
> Sebastian Haase
> UCSF, Sedat Lab
Dr Peter J Verveer
European Molecular Biology Laboratory
Tel. +49 6221 387 8245
Fax. +49 6221 397 8306
More information about the NumPy-Discussion