tlug: Re: linux cluster

["Manuel M. T. Chakravarty" <chak@example.com>]

Andrej Sali <sali@example.com> wrote,

> In comparison to single processor boards, do people find a big decrease 
> in performance when using two-processor boards, with large
> floating-point
> intensive programs that rely on large arrays (>> 512kB)?

I don't have any hard data for you, but if

* most of your computation is in-core (ie, not much disk
  access; you already mentioned that there is not much
  network traffic) and
* you have the typical scientific program that runs arrays
  up and down, then

your main bottleneck is CPU <-> memory bandwidth.  This
implies two things:

* Buy the biggest cache you can get (and write your
  application cache-friendly) and
* don't use two-processor boards.

The two processor will only end up competing for bus access,
instead of doing your float calculations.

Cheers,

Manuel

-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
   Manuel M. T. CHAKRAVARTY                      chak@example.com
   University of Tsukuba, Japan     http://www.score.is.tsukuba.ac.jp/~chak/
   SCORE Lab.                  Inst. of Information Sciences and Electronics
---------------------------------------------------------------
Next Nomikai: 20 November, 19:30 Tengu TokyoEkiMae 03-3275-3691
Next Meeting: 12 December, 12:30 Tokyo Station Yaesu central gate
---------------------------------------------------------------
Sponsor: PHT, makers of TurboLinux http://www.pht.co.jp