[CLUE-Tech] Best processor for the buck

[Ed Hill]

On Sun, 2004-10-17 at 10:09, Dale K. Hawkins wrote:
> We need a computer to handle some pretty hairy calculations.  We have
> run some tests and estimates and believe our current 1 GHz box will take
> nearly a week to finish the calculations.  And then we will want to make
> some modifications and re-run the calculations.  So, I need a speed
> daemon; a computer capable opening a can of whup-ass on some ugly math
> problems.
> 
> So I narrowed my decision to either a the P4 HT series or maybe a dual
> Opteron system.
> 
> I plan to run Debian (64-bit for the Opterons).
> 
> We have a computer guy with some suggestions, but his opinion of the
> 64bit systems is based on dealings with microsoft, vs. a natively
> compiled OS.
> 
> Anyhow, I'd appreciate any feedback.


Hi Dale,

Most (not all, but certainly most) scientific computing and "number
crunching" codes are memory bandwidth limited.  If you're thinking of
getting a dual-CPU system to speedup something that does a lot of number
crunching, then look into the dual Opterons.

As an example, our GFD code (MITgcm) gets a nearly-linear speedup
(relative to 1 CPU) when we run two simultaneous copies of it on a dual
Opteron system (eg. a Monte Carlo or Kalman filter analysis).  The exact
same setup will receive little improvement (in some cases, will actually
run *slower*) on a dual-P4 or dual-Xeon.  Why?  Dual Opterons can have
two memory controllers which doubles the total memory bandwidth while
current dual Intel chipsets do nothing to improve memory bandwidth when
you add the second CPU (actually, they tend to hurt it due to bus
contention).

Ed

-- 
Edward H. Hill III, PhD
office:  MIT Dept. of EAPS;  Rm 54-1424;  77 Massachusetts Ave.
             Cambridge, MA 02139-4307
emails:  eh3 at mit.edu                ed at eh3.com
URLs:    http://web.mit.edu/eh3/    http://eh3.com/
phone:   617-253-0098
fax:     617-253-4464