Re: confused about performance
Giacomo Mulas wrote:
> Two or three months, when I last compiled the latest version of the big
> quantum chemistry code (NWChem) I use (which spends a lot of time doing
> floating point linear algebra). The computer on which I tested is a
> relatively old Athlon 64 3500+, your mileage may vary on other machines.
> Oddly enough, an old version of GOTO was the fastest, followed closely by
> the optimised acml, then head to head the internal implementation provided
> by the quantum chemistry package and the (then) current GOTO, then atlas.
> Differences among all but atlas were measurable (i.e. reproducible) but
> very
> small, within 2%, atlas was ~10% slower. The Intel compilers produced much
> faster code than the gcc suite (both 3.4 and 4.1), despite running on AMD
> processors. This is a VERY specific test, of course, so I do not claim
> my conclusions should apply to anyone but me. On the other hand, they
> do apply rather well to my scenario by definition :)
I used to do this stuff (tune quantum chemistry codes) for a living.
Nostalgia ain't what it used to be. :) The only one I spent any great
length of time on was Gaussian (80 way back then) and it proved
extremely difficult to vectorize and parallelize, despite the efforts of
the developers to make it work well on such beasts as the CDC 205, which
had very deep pipelines. If NWChem is anything like that, I'm not
surprised the Intel compilers do a better job than GCC -- I don't think
GCC knows much about all the specifics of tweaking such things as
keeping data in caches, re-use, chip-level parallelism, etc. If NWChem
is open source, I'm sure someone will come along and profile/tweak it.
> I will obviously evaluate again atlas and new versions of the gcc suite
> if/when it's worth the effort. I look forward seeing an up-to-date version
> of atlas being included in debian. I would actually be very glad to be able
> to switch to a completely clean environment using gcc, since I currently
> have to keep around hosts of libraries compiled with different compilers
> and
> it's somewhat messy to maintain.
I wouldn't throw away that Intel compiler just yet. For that matter, I'd
give serious consideration to switching to a Core 2 Duo and a copy of
Intel's tuning tools ... they are quite good. Life's too short to wait
for calculations. :)
>
> Bye
> Giacomo
>
Reply to: