low cost parallel computing or supercomputers?

[Mike]

Hi all,

Could anybody point me to some good supercomputer or high performance
clusters with relatively low cost? I have wrote my first MPI program and it
is a coarse level parallelization, and I have to run many copies of it in
order to get back results in a few hours. Without using a large amount of
processors, I had been waiting for ages... I don't have a sponsor so I am on
my personal funding now and this is not for a commercial project... Thanks!

[manolinux]

Hi,

Maybe if your project or calculations are for scientific purposes, you
can join a networking grid.

Read about BOINC, for instance or any other kind of "volunteering"
computing.

http://boinc.berkeley.edu/

Grid is suitable for data parallelism, instead of paralellizing your
algorithms, you launch many copies of your program which work into a
different subset of data; and the underlying infraestructura deals with
launching the programs in the most suitable computer, relaunching if
there is some trouble, etc.

Regards!
Manuel.

Post by Mike
Hi all,
Could anybody point me to some good supercomputer or high performance
clusters with relatively low cost? I have wrote my first MPI program and it
is a coarse level parallelization, and I have to run many copies of it in
order to get back results in a few hours. Without using a large amount of
processors, I had been waiting for ages... I don't have a sponsor so I am on
my personal funding now and this is not for a commercial project... Thanks!

[Mike]

Post by manolinux
Hi,
Maybe if your project or calculations are for scientific purposes, you
can join a networking grid.
Read about BOINC, for instance or any other kind of "volunteering"
computing.
http://boinc.berkeley.edu/
Grid is suitable for data parallelism, instead of paralellizing your
algorithms, you launch many copies of your program which work into a
different subset of data; and the underlying infraestructura deals with
launching the programs in the most suitable computer, relaunching if
there is some trouble, etc.
Regards!
Manuel.

I read it. It's very cool. But I didnot find how to run my own big project
on it. Do you know? thankx

[manolinux]

Read this:
http://boinc.berkeley.edu/create_project.php
http://boinc.berkeley.edu/api.php

You have to do two main tasks:
* To settle a server which splits the works on clients
* Re-touch a little bit your apps programs, following an API, but think
less complicated to learn than MPI, and more suitable for data
parallelism (instead of algorithms).

And third, convince people to download your client program and start
donating CPU cycles for you ;)

Anyway, you can always mount this in intranet instead of through
internet ...

Much success !

Post by Mike

Post by manolinux
Hi,
Maybe if your project or calculations are for scientific purposes, you
can join a networking grid.
Read about BOINC, for instance or any other kind of "volunteering"
computing.
http://boinc.berkeley.edu/
Grid is suitable for data parallelism, instead of paralellizing your
algorithms, you launch many copies of your program which work into a
different subset of data; and the underlying infraestructura deals with
launching the programs in the most suitable computer, relaunching if
there is some trouble, etc.
Regards!
Manuel.

I read it. It's very cool. But I didnot find how to run my own big project
on it. Do you know? thankx

[Carlos Moreno]

Post by manolinux
Grid is suitable for data parallelism, instead of paralellizing your
algorithms, you launch many copies of your program which work into a
different subset of data [...]

Isn't this essentially paralellizing the algorithm?

I don't have much formal learning about parallel algorithms, but all
the instances I've seen rely precisely on splitting the data into
many chunks, launching one process on each CPU to deal with each
chunk, and then merge the results (an example of this is 3D CGI
rendering --- different sections of the image are completely
unrelated, and so they can be computed concurrently/simultaneously).

What am I missing?

Saludos,

Carlos
--

[manolinux]

Well, yes you're right.

But ok, think the person asking (or it is my interpretation) needs to
perform the same sequential operation on many independent sets of data
... so thats what i call "paralellism on data" ... instead of message
passing programming by MPI.

Its a discussion on which should be properly paralellized, the
algorithm, the data, or both? There are solutions for one extreme, the
other, or hybrid.

I only talked about BOINC because of my interpretation of the user
needs ... I mean, i perform some programming in MPI, but sometimes find
difficult to paralelize algorithms because they are inherently
sequential, and the solution to performance is something like grid.

But anyway ... your quote is right.
Regards !

Post by Carlos Moreno

Post by manolinux
Grid is suitable for data parallelism, instead of paralellizing your
algorithms, you launch many copies of your program which work into a
different subset of data [...]

Isn't this essentially paralellizing the algorithm?
I don't have much formal learning about parallel algorithms, but all
the instances I've seen rely precisely on splitting the data into
many chunks, launching one process on each CPU to deal with each
chunk, and then merge the results (an example of this is 3D CGI
rendering --- different sections of the image are completely
unrelated, and so they can be computed concurrently/simultaneously).
What am I missing?
Saludos,
Carlos
--