Setting up a scientific cluster with Centos 6.3

Issues related to applications and software problems
arshad
Posts: 40
Joined: 2006/10/05 17:36:39

Setting up a scientific cluster with Centos 6.3

Post by arshad » 2012/08/13 18:32:39

Can anyone please provide me with some directions ( docs, links , manuals etc. ) on how to setup a computational cluster
(for running c/c++, fortran simulations ) with CentOS 6.3?

I only know how to set up workstations and its maintainance. Never done any server setup.

Please treat me with kid gloves.

Thanks,

Arshad

Czarcasmo
Posts: 7
Joined: 2012/08/02 14:58:56
Location: Washington, DC

Re: Setting up a scientific cluster with Centos 6.3

Post by Czarcasmo » 2012/08/14 19:27:44

http://www.centos.org/docs/5/html/Cluster_Suite_Overview/ch.gfscs.cluster-overview-CSO.html

Czar.

User avatar
jlehtone
Posts: 4523
Joined: 2007/12/11 08:17:33
Location: Finland

Re: Setting up a scientific cluster with Centos 6.3

Post by jlehtone » 2012/08/15 10:46:06

Simple "computational cluster"? [url=https://computing.llnl.gov/linux/slurm/]SLURM[/url]
Not in default repository though.

C/C++/Fortran are programming languages and not really descriptive of the nature of the computations that you want to do. You can have threaded, MPI, or plain huge bunch of single-threaded binaries that you need to run regardless of what language you have used.

There are many ways to get one program to run in one/many remote computers. SLURM is one way to do it.

Another facets of clusters are how they are maintained, whether they share disk-storage, and what happens when one node has a fault.

DouglasRyder
Posts: 62
Joined: 2012/08/02 13:53:01
Location: UK

Re: Setting up a scientific cluster with Centos 6.3

Post by DouglasRyder » 2012/08/16 16:54:49

Documentation may not help at this stage. You will probably learn much faster by building yourself a test system (perhaps using your existing workstations).

Can you give us a clue what sort of simulations you need to run?

arshad
Posts: 40
Joined: 2006/10/05 17:36:39

Re: Setting up a scientific cluster with Centos 6.3

Post by arshad » 2012/08/16 18:09:27

Thank you everybody.

And to answer DouglasRyder:

I have a bunch of workstations ( 8 of them now ) - all of which have2 quad core xeons with two 0.5 TB HDD and 8GB rams.
I run mostly single threaded monte carlo simulations of esoteric quantum field theories while my colleagues
run molecular dynamics simulations of liquids.

First Problem - what extra bunch of group of softwares do I install on top of the workstation class install.
Alternately, I can wipe off everything of these macines and reinstall from scratch.

If I try to follow the second route, what should be my modus operandi? I am assuming that the cluster should be high performance one ( not high-availability , or should
it be both ? ) .

Best wishes,

Arshad

DouglasRyder
Posts: 62
Joined: 2012/08/02 13:53:01
Location: UK

Re: Setting up a scientific cluster with Centos 6.3

Post by DouglasRyder » 2012/08/16 19:57:06

OK. I'm not familiar with the physics. What I was trying to find out is how the simulations run from an IT point of view. One of your applications is single threaded, so that should be easy to deal with.

What about the other one? Is it a single threaded or SMP solver that runs on a single machine, or is it a DMP/MPI solver that runs across multiple machines?

Could you please clarify what you are asking?

Are you planning to install a bunch of dedicated servers to do the number crunching, or are just looking for a job scheduler to manage your existing workstations?

Apologies for the questions, but I can't help until I understand what you are asking

Regards

arshad
Posts: 40
Joined: 2006/10/05 17:36:39

Re: Setting up a scientific cluster with Centos 6.3

Post by arshad » 2012/08/17 15:43:50

[quote]
DouglasRyder wrote:
OK. I'm not familiar with the physics. What I was trying to find out is how the simulations run from an IT point of view. One of your applications is single threaded, so that should be easy to deal with.

What about the other one? Is it a single threaded or SMP solver that runs on a single machine, or is it a DMP/MPI solver that runs across multiple machines?

Could you please clarify what you are asking?

Are you planning to install a bunch of dedicated servers to do the number crunching, or are just looking for a job scheduler to manage your existing workstations?

Apologies for the questions, but I can't help until I understand what you are asking

Regards[/quote]

I know the naive difference between DMP and SMP but not the details. Can't one have both?

The original codes were written for single threaded machine. We wish to port them to the newer
machines, We would like to use all of them and all the cores - but it must allow for several different jobs submitted by
different people to be run simultaneously.

Thanks for your patience,

Arshad

arshad
Posts: 40
Joined: 2006/10/05 17:36:39

Re: Setting up a scientific cluster with Centos 6.3

Post by arshad » 2012/08/18 04:43:29

We plan to use one as a master node,
the rest as slaves. Does it make it an SMP system?

DouglasRyder
Posts: 62
Joined: 2012/08/02 13:53:01
Location: UK

Re: Setting up a scientific cluster with Centos 6.3

Post by DouglasRyder » 2012/08/18 11:55:25

A DMP solver can run across multiple machines, each of which talks to its peers via an MPI comms layer. You want a dedicated HPC cluster for this.

An SMP solver runs on a single shared memory machine, so its easier to deal with. You just need a job scheduler that sends your jobs to a bunch of slave nodes. The slave nodes should ideally be dedicated servers, but you can use workstations if necessary.

You will need a job scheduler. The choices are LSF, TorquePBS, PBS-Pro, Sun Grid Engine, and Slurm. All are widely used, and you won't find anybody that can tell you which is the best (it takes to long to test them all).

You will also need an NFS server. Most job schedulers require that every user has the ability to run a ssh command on every slave node without a password. The easiest way of configuring this is to put everybody's home directory on an NFS server. This ensures that the public/private keys in $HOME/.ssh are available on every machine.

I'm confused by your last comment about a master node. You don't need one if you are running the jobs on your workstations. Please explain the layout of your hardware:

What machines are you physically sat in front of when you submit a job?
Are these Linux or Windows?
What machines do you intend to use for the number crunching?

arshad
Posts: 40
Joined: 2006/10/05 17:36:39

Re: Setting up a scientific cluster with Centos 6.3

Post by arshad » 2012/08/19 08:06:37

[quote]
DouglasRyder wrote:
A DMP solver can run across multiple machines, each of which talks to its peers via an MPI comms layer. You want a dedicated HPC cluster for this.

An SMP solver runs on a single shared memory machine, so its easier to deal with. You just need a job scheduler that sends your jobs to a bunch of slave nodes. The slave nodes should ideally be dedicated servers, but you can use workstations if necessary.

You will need a job scheduler. The choices are LSF, TorquePBS, PBS-Pro, Sun Grid Engine, and Slurm. All are widely used, and you won't find anybody that can tell you which is the best (it takes to long to test them all).

You will also need an NFS server. Most job schedulers require that every user has the ability to run a ssh command on every slave node without a password. The easiest way of configuring this is to put everybody's home directory on an NFS server. This ensures that the public/private keys in $HOME/.ssh are available on every machine.

I'm confused by your last comment about a master node. You don't need one if you are running the jobs on your workstations. Please explain the layout of your hardware:

What machines are you physically sat in front of when you submit a job?
Are these Linux or Windows?
What machines do you intend to use for the number crunching?[/quote]

We don't want direct access to the cluster. The master node is where the jobs will be submitted from outside machines.
Yes we want to run linux only ( else I wouldn't have posted on Centos forum :-D ).
And the crunching will be done at these workstation/servers which I described earlier.

Do you need more information?

I am thankful for your patience.

Arshad

Post Reply