[ILUG] Advice much appreciated.

[Ronan Cunniffe]

On Wed, 16 Feb 2005, Bernard-Joseph Roche wrote:

> Hi,
>
> I am looking for a new box (who isn't?) to run some number crunching
> on. I was thinking of getting the following bundle :
>
> http://www.komplett.ie/k/ki.asp?sku=111924&cks=PRL
>
> My query goes out to people using a system similar to the above. How
> have they found it? Any issues? I plan to run FC364-bit initially but
> may drop back to 32-bit if i hit problems. I am open to other distros
> where people have had good experiences with 64bit. I choose this
> bundle as it is cheaper than most P4 w/HT on the market.
>
> So any thoughts would be appreciated,

If you buy a mobo with a built in (shared memory) graphics card, beg,
borrow, or steal a crappy PCI card and disable the onboard one.  Otherwise
the graphics chip is consuming memory bandwidth.

Other thoughts:  (I'm a sysadmin for numerical computation people, I get
seconds)

The OS is irrelevant
---------------------
   A number crunching box is basically a fast calculator with an OS
somewhere vaguely nearby for when we need it.  Apart from file I/O and
swapping, the OS doesn't enter the picture (and ideally neither of
those things should happen much).

   The only reason to use a 64-bit OS for number-crunching is if the
application specifically takes advantage of the 64-bitness.  If you don't
know that it does, it probably doesn't.  If the app's working dataset is
less than 2GB, it probably doesn't.  If it does, it probably doesn't
matter much.  Just say N.


HT is (probably) irrelevant
---------------------------
Hyperthreading lets two parallel threads from the same process (i.e sharing an
address space - the cache isn't partitioned) be on chip at once.  Most
number-crunching stuff is single threaded, and even the multi-threaded
stuff is usually "computation-plus-occasional-batch-I/O-we-stuck-in-a-
separate-thread-just-cos", which makes negligible difference unless it's
really badly written.


CPU cache size may be seriously important
-------------------------------------
I had someone with a problem with a 1.5MB working set.  We ran on a
little laptop (P-M 1.4GHz, 2MB cache, 400MHz FSB) and a server CPU
(Xeon 3.066GHz, 512MB cache, 533FSB).  Laptop won comfortably (~30 hours
vs. ~35) despite being outclocked 2:1.


If CPU cache size is not important, then memory size and speed will be
----------------------------------------------------------------------
With an interesting twist: if slower speed grades of RAM are significantly
cheaper, it *might* make sense to get an older machine with lots of DIMM
slots and load it to the gills with cheaper RAM - lifting more of your
problem off disk.  (This is particularly good if you have overclocker
friends you can scrounge older types of RAM off!)


Stability is more important than performance
--------------------------------------------
If there was a 10% performance improvement running a 64-bit kernel over
a 32-bit kernel, I'd run the 32-bit kernel.  If there was more than a 10%
difference, I'd get very paranoid about the compiler and very careful to
cross-check results between 32-bit and 64-bit runs.


Apart from all that - without a magic calculator that gives the
performance of the specific application for every box you spec, it's
impossible to identify the absolutely-perfect-for-my-application machine.

It *is* possible (and fairly straightforward) to spec a machine that will
get you reasonably close to that theoretical peak.  Buy that machine and
be happy.

Hope this is of some use,

Ronan