[ILUG] Getting started with Beowulf

[Cian Davis]

Hi Robert,
You mentioned fluid dynamics so I assumed CFD.

The whole question of benefit of parallel computing depends on how well
the problem can be distributed to other systems. The more discrete
chunks the calculation can be broken up into and the less communication
that is needed between the chunks, the better the parallel performance
will be. However, optimising either of these is not a trivial matter.

>From experience, you need a very large problem to take advantage of more
than 8 cores or processors. My research involved CFD problems with more
than 2 millions cells (To give you an idea of computations done, each
cell required solution of 6 differential equations over 1,000
iterations). With 8 processors, the speedup relative to a single
processor was about only about 3.5 (though memory requirements may
require multi-processor solution).

The point being, before you go rewriting code, setting up clusters and
all the rest, you need to look and see how well you think it will break
up, how long it will take you to rewrite, setup, test etc. and see if
that is greater than the speed improvements you are likely to get. This
is especially true if this is part of research you're doing (if so,
there would also be a risk that you deviate from your research aims).
You might well be better investigating lots of parameters be solving
using a large number of stand-alone machines.

There was a very good conference last year in UCD about writing MPI
software. The tutorials were particularly good. I was hoping that the
papers would be on the website (http://pvmmpi08.ucd.ie/) but not so. I
will contact you off-list with a contact. This years conference seems to
be in Finland: http://www.csc.fi/english/pages/pvmmpi09/

Regards,
Cian

Robert Grundulis wrote:
> Hi Cian,
>
> My experience with clustering and HPC is limited, but I'm learning fast.
>
> I've read into MPI and I can already see how I could rewrite much of my code
> to take advantage of it.
>
> As for CFD, I'm not sure how big a part it plays in all this. The
> calculations I am being provided with are unique to the field of naval
> architecture. I'm sure CFD is factored in to them, buit for the most part
> 80% of the formulae I've been given are simply serial argorithms, repeated
> many times and I've not come across talk of cells.
>
> I should also point out I've spent the last week trying to get in to
> Stroud's Engineering Mathematics book and I'm still learning about
> enginnering math. It's a lot different than writing invoicing software!
>
> Robert
>
> 2009/6/24 Cian Davis <cian.davis at skynet.ie>
>
>   
>> Hi Robert,
>> During my Ph.D. (on CFD), I installed an maintained a small Beowulf
>> cluster for our group and then spend a postdoc year installing and
>> configing a larger system in UL (http://photos.killminus9.net/cluster/6/).
>>
>> I don't know how much experience you have with CFD, Beowulf/HPC or MPI
>> so apologies if I state some stuff that's obvious!
>>
>> First off is that HPC will not always speed up the solution. You need a
>> certain amount of cells (in CFD) per computational core and then the
>> efficiency of programming, the MPI used, the interconnect, data volumes,
>> hardware used etc. all have a significant bearing on how much faster the
>> problem will run.
>>
>> If you let me know what software you're using, how many cells and a bit
>> more on the hardware you have available (e.g. amount of RAM, hard disks,
>> network interconnect), I'd be more than happy to give you some advice on
>> the best way to proceed.
>>
>> Regards,
>> Cian Davis
>>
>> Robert Grundulis wrote:
>>     
>>> Hello all,
>>>
>>> Can anyone offer me suggestions on good books or resources that would
>>>       
>> help
>>     
>>> get me up to speed with Beowulf clustering?
>>>
>>> I've written an application that does calculations on fluid dynamics,
>>> propellers and hull forms (boaty stuff) and the intention is to move it
>>>       
>> on
>>     
>>> to a cluster to speed it up. At the moment one calculation takes about 3
>>> days on a dual core Xeon, so there are obvious benefits to moving it on
>>>       
>> to a
>>     
>>> cluster.
>>>
>>> I have free access to dozens of old university PCs and servers (all P4
>>>       
>> and
>>     
>>> adbove) and the plan is to cluster them.
>>>
>>> The thing is I've never worked with clusters before and I don't have a
>>>       
>> clue
>>     
>>> where to start. Can anyone help point me in the right direction?
>>>
>>>
>>>       
>>     
>
>
>