Re: confused about performance
On Thu, Jun 14, 2007 at 04:08:54PM +0200, Leopold Palomo-Avellaneda wrote:
> really, reading you makes me doubt about the whole port. How many apps do we
> have in the debian pool that can win some kind of performance?
If nothing else, running a 64bit kernel with 32bit user space is an
option. That gives you support for more than 4GB ram without the
penalty of PAE, but leaves user space limited to 3GB per process. I run
one system that way and it runs great. Systems with that much ram are
starting to be very common.
And as I said anything using floating point will benefit quite a bit.
gcc will hopefully also get better at optimizing for the amd64
architecture over time.
Now I guess the question is whether having more than 3GB ram per process
is useful? Well for many databases it could be. Maybe for image
editing programs it will be. Some applications can be written simpler
if you can simply memory map a large file and work on it that way rather
than using fseek/fread/fwrite. Being able to memory map files larger
than 2GB certainly has benefits for video editing and such.
I guess the real question then is: Does the benefits for those programs
give enough to warrent a tiny loss on many smaller utilities? Running
mixed 32/64bit on one system is just a pain, so in my book the trade off
is worth it. I don't think I will notice the 1 or 2% potential loss due
to the larger pointers, while I probably will notice the benefits of
faster floating point (perl uses lots of floating point by the way) and
direct access to more ram in larger programs.
Now I just tried a small stupid test here:
I have a zip file that is 61592222 bytes. I tried running
'time tar cvzf test.tgz my.zip' on both 32 and 64 bit on an Athlon 64 3500+
running a 64bit kernel with 1GB ram.
According to time, it takes 0m3.776s of user time (so not counting
system calls) on 64bit, and 0m4.556s on 32bit. That indicates tar/gzip
runs 20% faster on 64bit than 32bit at least when the kernel is 64bit
and run on the same machine. I ran them a few times each to make sure
caching and such didn't affect them. Maybe your test method wasn't
correct or the two systems aren't as identical as you think.
--
Len Sorensen
Reply to: