Science > Physics > building a small-college cluster for computational physics
| Topic: |
Science > Physics |
| User: |
"nt moore" |
| Date: |
14 Feb 2006 11:49:17 AM |
| Object: |
building a small-college cluster for computational physics |
Suppose you're a professor at a small college with 10-20K startup funds
and you want to build a cluster for your major students to use to do
computational physics work. What setup would you propose?
requirements:
- cluster runs linux
- needs to be useful for (pedagogically oriented) research for 5 years
- need to have MPI available
- need to have a job scheduling system
Right now, the most interesting thoughts I've had are along the lines
of having a several (5-7) mac mini's available as "workstations" and an
equal number of rackmount servers, but the apple xserves are expensive
and starting with a heterogenous cluster seems like a headache.
.
|
|
| User: "" |
|
| Title: Re: building a small-college cluster for computational physics |
14 Feb 2006 12:50:28 PM |
|
|
Why not use Citrix for this configuration? With the Citrix set-up, you
could use very low-end and cheap hardware as workstations and a very
robust server cluster for number crunching.
.
|
|
|
| User: "physicsboy" |
|
| Title: Re: building a small-college cluster for computational physics |
14 Feb 2006 12:55:28 PM |
|
|
wrote:
Why not use Citrix for this configuration? With the Citrix set-up, you
could use very low-end and cheap hardware as workstations and a very
robust server cluster for number crunching.
Citrix is a good suggestion. However, closely re-examine the
requirements that he specified. Must run Linux. While there is a UNIX
version of Citrix Presentation Server, no Linux version of the server
component has been released yet.
Linux Terminal Server would be an alternative to this problem.
.
|
|
|
|
|
| User: "physicsboy" |
|
| Title: Re: building a small-college cluster for computational physics |
14 Feb 2006 01:09:04 PM |
|
|
nt moore wrote:
Suppose you're a professor at a small college with 10-20K startup funds
and you want to build a cluster for your major students to use to do
computational physics work. What setup would you propose?
requirements:
- cluster runs linux
- needs to be useful for (pedagogically oriented) research for 5 years
- need to have MPI available
- need to have a job scheduling system
Right now, the most interesting thoughts I've had are along the lines
of having a several (5-7) mac mini's available as "workstations" and an
equal number of rackmount servers, but the apple xserves are expensive
and starting with a heterogenous cluster seems like a headache.
I have created a similar setup using Redhat Fedora Core 3 and Linux
Terminal Server software. All the client workstations are very low-end
have no disk drive (they load from a network drive on the Terminal
Server). Our setup is fairly unexpensive and it is enough for almost
any type of calculation and data analysis.
.
|
|
|
| User: "" |
|
| Title: Re: building a small-college cluster for computational physics |
14 Feb 2006 03:21:27 PM |
|
|
You could also use Google's approach in that they use a whole lot of
low profile computers to add more and more computing power to the
cluster. This approach is called distributed computing and you can use
any computer 3-4 years old that someone donated to you. The data would
be stored in a central network location but the processing power comes
from the addition of all the computers. The only downside to this
approach is the space to put all those computers, the noise that will
result from all those computers running at once and the power
consumption but I have seen it successfully implemented at many
institutions.
.
|
|
|
| User: "" |
|
| Title: Re: building a small-college cluster for computational physics |
15 Feb 2006 09:07:54 AM |
|
|
How is the project coming along?
.
|
|
|
|
|
|
| User: "Puppet_Sock" |
|
| Title: Re: building a small-college cluster for computational physics |
14 Feb 2006 01:01:19 PM |
|
|
nt moore wrote:
Suppose you're a professor at a small college with 10-20K startup funds
and you want to build a cluster for your major students to use to do
computational physics work. What setup would you propose?
requirements:
- cluster runs linux
- needs to be useful for (pedagogically oriented) research for 5 years
- need to have MPI available
- need to have a job scheduling system
Right now, the most interesting thoughts I've had are along the lines
of having a several (5-7) mac mini's available as "workstations" and an
equal number of rackmount servers, but the apple xserves are expensive
and starting with a heterogenous cluster seems like a headache.
Why is linux a requirement?
MPI is available for nearly any platform you might consider.
Any system that has students attached to it already has a
"job scheduling system." It's called "Hey, Joe. Be sure to
start that job when the first job finishes." Though pretty
much any OS you want to point at has some kind of
job scheduler.
You say you have a small budget, then you whine that
"apple xservers are expensive."
Why would you even consider Macs? If you really must have
linux, get Intel boxes and pile them up. Load up linux. Have
one LAN hard drive. Get a switcher box that lets you share
one monitor, keyboard, and mouse among a bunch of
CPUs. That will save you even more money as you only
need to buy as many monitors as you will have actual
people working on the system at any given time. You can
get cheapo PCs these days, particularly if you don't insist
on getting "bleeding edge" speed. And it's pretty easy to
swap in more CPUs, or faster ones, as funds become available.
A heterogeneous cluster is not necessarily a problem. If the
platforms are binary compatible it's almost transparent.
If the platforms are such that they use the same format
for representing data (floats, integers, strings, etc.) then
it is still pretty close to transparent.
Socks
.
|
|
|
|
|
Related Articles |
|
|
