[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Promoting i386 version over x86_64?

Hash: SHA1

King InuYasha wrote:

> Except, that could be false advertising. In most cases, where CPU
> computation is not used heavily, 64-bit is actually SLOWER than the
> 32-bit counterpart. Optimizations are narrowing the gap, but it still
> remains true. 

On ppc versus PPC64, sparc vs. sparc64, and possibly other architectures
that may be true, however on x86 vs. x86_64 arch as a whole it is not
generally the case, in particular because the x86_64 arch has double the
number of available registers for gcc to play with.

Whenever I have done any performance testing comparisons between x86 and
x86_64, all I/O bound processes tend to come out with very similar
results, however the more CPU bound a task is, the more likely the app
is to have up to a 30% performance gain depending on various factors.
gcc for example tends to build much faster on x86_64 than on x86.

The only cases where I've personally seen an x86_64 built application
perform poorly compared to the i386 built version of the app on the same
system, when investigated - turned out to be that the source code of the
application had x86 specific assembly language which got used on the x86
build, and much slower C fallbacks when used on x86_64 (and other
arches).  There probably aren't a lot of packages in the distribution
that contain x86-only hand crafted assembler which end up using C
fallbacks on x86_64, but it is one possibility.

What applications are you aware of which run slower on x86_64 than on
x86 on the same system?  It would be interesting to investigate
whichever ones you've discovered to find out why they are slower as it
shouldn't generally occur, although I'd suspect it would be a case of
fast path x86 specific optimizations with a slow path for non-x86 as
mentioned above.

Just curious what you may have observed differs.

- --
Mike A. Harris                   Website: http://mharris.ca
Google Wave: mike.andrew.harris - at - googlewave.com
https://identi.ca/mharris | https://twitter.com/mikeaharris
Version: GnuPG v1.4.5 (GNU/Linux)


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]