LAM: Cluster config / node setup

Greetings,

My name is Romulo Cholewa, and Iīm currently working on
designing an HPC
solution that will run LAM/MPI (early stage).

I have some understanding about the basics, but the last
time Iīve
technically played with HPC was with MOSIX years ago. Now
Iīm working on the
pre-sales side, and couldnīt find any documentation about
potential system
designs.

Fact is, we need to design a system capable of good
performance, future
scalability and some cost in mind (but not that much). Time
is also an
issue.

The main idea is to build the HPC with IBM BladeCenter. We
want to start
with 6 nodes / blades, each one with 4 to 6 GB of RAM, 2 *
Intel QC 2,66 GHz
(3550). This way we should prove the usefulness of the
concept and attain a
wonderful density, if we need more nodes.

Starting questions

. Should we forget general redundancy in favor of
connectivity / bandwidth ?
. Should we start with infiniband / myrinet and local disks,
or blades with
2 * 1 Gbps Ethernet and remote storage (SAN), without local
disks ?

Point in mind

If we choose local disks and we have to increase the number
of nodes,
storage management may become harder. Booting all
blades/nodes from a
central SAN storage might easy things up. We canīt
technically have a blade
with an infiniband module and an HBA at the same time atm,
but we can use
the infiniband for node communication and SAN access. Would
it be the way to
go ?

I think these are rather newbie questions, so if anyone have
any URLs
pointing to relevant info, it would be great.

Thanks in advance,

Romulo M. Cholewa
Info & PGP:  [http://www.rmc.eti.br]
Disclaimers: [http://www.rmc.eti.br#ema
il]
EMail/IM:    [rmc at rmc.eti.br]




_______________________________________________
This list is archived at http://www.l
am-mpi.org/MailArchives/lam/

Romulo M. Cholewa wrote:

> 
> The main idea is to build the HPC with IBM BladeCenter.
We want to start
> with 6 nodes / blades, each one with 4 to 6 GB of RAM,
2 * Intel QC 2,66 GHz
> (3550). 
> .. Should we start with infiniband / myrinet and local
disks, or blades with
> 2 * 1 Gbps Ethernet and remote storage (SAN), without
local disks ?
>  if anyone have any URLs
> pointing to relevant info, it would be great.

InfiniBand is well worth while, unless your applications are
limited to
those with little dependency on MPI performance.  Openmpi,
Intel MPI, HP
MPI are more widely supported bases for IB support.
An adequate parallel file system seems overkill for a small
cluster.

http://softwarecommunity.intel.com/articles/eng/1311.htm


_______________________________________________
This list is archived at http://www.l
am-mpi.org/MailArchives/lam/

On Thu, 11 Oct 2007, Romulo M. Cholewa wrote:

> Now Iīm working on the pre-sales side, and couldnīt
find any 
> documentation about potential system designs.

I think that the beowulf list is more appropriate for this
type of 
questions. Especially as I see your questions with very
little 
relevance to LAM/MPI...

> . Should we forget general redundancy in favor of
connectivity / bandwidth ?

Have you read anything about redundancy support in LAM/MPI ?
There is 
very little, if any... In any case, you should first define
your 
requirements for redundancy and then see if/what support if
available.

> . Should we start with infiniband / myrinet and local
disks, or blades with
> 2 * 1 Gbps Ethernet and remote storage (SAN), without
local disks ?

Have you checked which Infiniband and Myrinet
drivers/libraries are 
supported by LAM/MPI before asking yourself whether to add
or not the 
hardware to your configuration ?

LAM/MPI cares very little about storage; each daemon needs
access to 
/tmp (or some other place which you can choose) but it's
only for 
keeping some state, nothing I/O intensive.

> I think these are rather newbie questions, so if anyone
have any 
> URLs pointing to relevant info, it would be great.

Well, http://www.lam-mpi.org/fa
q/ would be a great place to start 

FWIW, you should also look at Open MPI: http://www.open-mpi.org/


--
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches
Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg,
GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.CostescuIWR.Uni-Heidelberg.De
_______________________________________________
This list is archived at http://www.l
am-mpi.org/MailArchives/lam/