[wellylug] distro performance

Brent Wood pcreso at pcreso.com
Sat Aug 26 12:32:48 NZST 2006


I figured this might be of interest to a few members. I think the conclusion to
be drawn is that small changes can make big differences, but depending on your
application you might get very different results.

We have developed a stock assessment model for fisheries work. Floating point
intensive, reasonable memory I/O & very little disk I/O. If you really want to
know about it see http://www.niwascience.co.nz/ncfa/tools/casal

On a mid range PC (say P4 2400/1Gb memory) real world simulations often take up
to a week, & several dozen are often needed to be run.

I've mentioned before how running this under Linux is much faster than Windows.
Possibly a fuction of mingw/gcc efficiency, but as pragmatists, we'll go with
Linux, why is not really an issue. We just want it to run fast.

We have also found AMD 64 cpu's with 64bit Linux using a late 2.6 kernel & gcc
v4.x are faster than Intel cpu's, earlier kernels & gcc 3.x. (not enough info
to really compare 4.0 with 4.1, but I'd guess 4.1 is potentially faster) Nvidia
chipset probably faster than VIA, but nothing really conclusive on this one.


We are putting together a mini cluster, all A64 4600+ dual core cpu's with 4Gb
RAM in each. One queen bee, 6 diskless drones booting off it, & using a
clustering solution (probably CONDOR) to assign jobs to each core.


This is the interesting bit. As of last thing Friday....

We stuck Fedora core 5 on it, compiled CASAL & ran a small benchmarking model.
A reasonably quick 3.5 minutes.

Then threw SLED10 on it & did the same compile & benchmark. 2.0 minutes.

Dunno why, but we're happy :-) That's the fastest this benchmark has ever run.
Novell claim to have optimised SLED for computational work, but this seems a
little too good.

What has become apparent throughout this exercise, is that cpu's, chipsets,
OS's & distro's can have quite large effects on performance, & depending on
your benchmark, you can get very different results.

Users of a browser or office suite won't see the same difference, but for a
specialised app, you can get some very significant improvements with a few
hardware/software changes. 


Cheers,

  Brent





More information about the wellylug mailing list