Performance measurements (was: python on the smalltalk VM)

[Andrew Dalke]

Douglas Alan wrote:
>All this should come as no shock -- 100 times slower for an
>interpreted language is just par for a good interpreter.  Before the
>spiffy Self compiler, state of the art was 10x slower than C for a
>really bitchin' highly optimized Smalltalk compiler.

I have a package called PyDaylight which puts a deep Python
layer (with OO, exceptions, iterators, etc) on top of a C library
for chemical informatics.  I've done a lot of timing tests on
the library to see how the Python code compares to morally
equivalent C code.

In most cases it's about a factor of ten.  The fastest case
is a 50% slowdown for code which is mostly in the extension -
and it should be faster now that Python has xreadlines.  The
slowest case was a factor of 40 for code which did a lot
of array lookup and used the __getattr__ hooks

My timing tests have put Python code anywhere from 50% slower
than C code (when most of the work was done in a C extension)
to 40x.  This slowest case was for code that did a lot of list
lookups instead of pointer increments, and used the object
layer heavily, which is translates attribute lookups to the
C level via getattr calls.

Most code is O(10x) slower.  Of course, to balance it out
in most cases the code itself is an order of magnitude fewer
lines of code and an order of magnitude easier to read.

So I don't agree with your estimate of 100x slower, at least
for the domain of programs I've investigated.

                    Andrew
                    dalke at acm.org