opteron VS Itanium 2

Thomas Alrutz Thomas.Alrutz at dlr.de
Wed Oct 29 11:15:48 EST 2003

Hi Gabriele,

we have bought a similar Linux Cluster (16 nodes) you are lokking for 
with the smallest dual Opteron 240 (1.4 GHz) and two Gigabit networks
(one for communications (MPI) and one for nfs).

> Dear all, 
>       we are planning to build up a new cluster (16 nodes) before this
> year's end; we are evaluating different proposals from machine sellers,
> but the main doubt we have at this moment is whether choosing an Itanium
> 2 architecture or an AMD Opteron one. 
> I know that ther's had already been on this list a debate on such a
> topic, but maybe some of you has some new experience to tell about. 

The nodes have all 2 GB RAM (4*512 MB DDR333 REG), 2 Gigabit NICs 
(Broadcom onboard) and a Harddisk. The board we had choosen was the 
Rioworks HDAMA. I know it is not cheap, but it is stable and 
performances well with the SUSE/United Linux Enterprise Edition.

> There is a wild bunch of benchmarks on these machines, but we fear that
> these are somewhat misleading and are not designed to test CPU's for
> intense scientific computing. The code we want to run on these machines
> is basically a home-made code, not fully optimized, which allocates
> around 500 Mb of RAM per node. Communication between nodes is a quite
> rare event and does not affect much computation time. In the past we had
> a very nice experience using Alpha CPU's which performed very well.

We have done some benchmarking with our TAU-Code (unstructured finite 
volume CFD-code, in multigrid), which hangs extremly on the memory 
bandwith and latency. Therefore we tested 4 different architectures:

1. AMD Athlon MP 1.8 GHz   FSB 133 MHZ - with gcc3.2 in 32 Bit
2. Intel Xeon 2.66 GHz     FSB 133 MHZ - with icc7   in 32 bit
3. Intel Itanium2  1.0 GHz FSB 100 MHZ - with ecc6   in 64 Bit
4. AMD Opteron 240 1.4 GHz FSB 155 MHZ - with gcc3.2 in 64 Bit

For the benchmark we used a "real life" example (aircraft configuration 
with wing, body and engine - approx. 2 million grid points) which 
desires 1.3 GB to 1.7 GB for the job (1 process)
We have performed 30 iterations (Navier Stokes calculation - Spalart 
Allmares - central scheme - multigrid cycle) and taken the total 
(Wallclock) time.

> To sum up, the question is: is the Itanium2 worth the price difference
> or is the Opteron the best choice?

To answer your question take a look on the following chart :

All times in seconds for 1 cpu on the node in use

1. AMD Athlon MP 1.8 GHz   - 30 iter. = 3642.4 sec.
2. Intel Xeon 2.66 GHz     - 30 iter. = 2151.4 sec. <- fastest
3. Intel Itanium2  1.0 GHz - 30 iter. = 3571.8 sec.
4. AMD Opteron 240 1.4 GHz - 30 iter. = 2256.5 sec.

and 2 cpu on the node in use (2 process via MPI)

1. AMD Athlon MP 1.8 GHz   - 30 iter. = 2076.1 sec.
2. Intel Xeon 2.66 GHz     - 30 iter. = 1447.8 sec
3. Intel Itanium2  1.0 GHz - 30 iter. = 1842.8 sec.
4. AMD Opteron 240 1.4 GHz - 30 iter. = 1159.5 sec. <-- fastest

So here you can see why we had to choose an Opteron based node to build 
up the cluster.
The price/performance ratio for the Opteron machine is verry good 
compared to the itanium2 machines.
And the Xeons are not so much cheaper....

  __/|__ | Dipl.-Math. Thomas Alrutz
/_/_/_/ | DLR Institut fuer Aerodynamik und Stroemungstechnik
   |/    | Numerische Verfahren
     DLR | Bunsenstr. 10
         | D-37073 Goettingen/Germany

Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

More information about the Beowulf mailing list