Fortran vs Python - Newbie Question

[sturlamolden]

On Mar 26, 7:13 pm, "Noma... at gmail.com" <Noma... at gmail.com> wrote:

> Thanks you all for giving a little insight into what Python can
> actually do, I think I've read enough to convince me that Python is
> generally a very flexible, fast, powerful language that can be used in
> a wide variety of applications instead of focusing on numerical
> functions like fortran does.

I wouldn't go as far as to say Python is a 'fast' language. But it's
certainly fun (particularly that), powerful, readable, and useful,
even for numerical computing.

Speed in numerical computing is a strange beast. First, a small
bottleneck in your code is likely to dominate the execution time. But
very often it is impossible to guess where it is, common sense usually
doesn't help, for some strange reason. You can use the Python profiler
to identify the real bottlenecks, but you may be surprised when you
find where they are. These bottlenecks are the only portion of the
code that can benefit from being moved to Fortran or C. The
bottlenecks will dominate regardless of what you do to the rest of you
code.

You may gain nothing from moving the bottlenecks to C or Fortran.
Don't let that surprise you. Python may be doing the work as fast as
it can be done. It may actually be the case that moving this code to C
or Fortran causes a slowdown. Python may be calling hand tuned C,
assembly or Fortran code that you cannot match. You can hand code a
sorting routine in C, but will you be able to beat Python's 'timsort'?
Probably not! Timsort is one of the fastest and stable sorting
routines known to man. It has taken years to develop and fine tune,
and you simply cannot beat that, even if you spend hours on end
optimizing your code. The same holds for the hashing routines in
Pythons dictionaries. You can write a hashtable in C, but I doubt it
can match the speed of Python's dictionaries. And then there is files,
strings, network sockets and protocols, access to databases,
encryption, etc. Python's libraries are tuned by people who knew what
they were doing and had the time to do the fiddling. Most likely you
don't (well, I certainly do not). Finally, you will most likely use
numerical libraries like LAPACK and BLAS. These are written in Fortran
or C anyway. Whether you call the libraries from Fortran or Python may
not matter very much! In the end, the code that can really benefit
from being moved to C or Fortran is rare to come by.

Python is a very high-level language. That means there are certain
things that put constraint on the attained speed. Most importantly:
keep the number of interpreter evals as scarce as possible. If you
make a for loop, the interpreter may evaluate the lines within the
loop several times. Languages like C, C++ and Java really teach you
bad habits when it comes to an interpreted language like Python. In
these languages, loops are almost free. In Python, they may be very
expensive as the interpreter is invoked multiple times. But if you can
'vectorize' the loop statement into a one-liner (or a few lines) using
slicing, list comprehensions or functional programming (e.g. the
lambda, map and filter intrinsics), the overhead will be very small.
That is the secret to getting high-performance from a language like
Python. If you can get away from the C++ habit of using for loops for
any kind of iteration, Python may not feel sluggish at all.

In the end, the main determinant of performance in numerical tasks is
really big-O, not the language you prefer to use. Choice of algorithm
is fare more important than the Python vs. Fortran issue!