[Carpet] PsiKadelia and Carpet

[Erik Schnetter]

On Nov 14, 2007, at 01:09:42, Hee Il Kim wrote:

> Thanks Erik,
>
> - I think the problem was caused by memory excess. I took grid  
> sizes too large. I thought it is safe because the run lasted more  
> than 100 timesteps but suddenly it used up a node memory. I'm not  
> sure whether this is a memory leak. But I don't see any drastic  
> memory increase for smaller grid sizes even though it is not  
> observed for a long time. I will ask you about more details if the  
> problem turns out to be a serious one.

This could also be caused by the HDF5 library.  You can try using  
version 1.8, which should use less memory that version 1.6.

> - I didn't try to make an account for LSU HPC yet. If I make the  
> account, then can I get 10,000 SUs as other users or .. ? Which  
> machines can I use for Cactus run?

You would probably get a startup allocation of 10 kSU.  For larger  
allocations you would need to write an application.

> - It's not related to Carpet code but I'd like to hear your advice.  
> I will make( or buy) a new cluster for Cactus/Whisky run. It will  
> be used for debugging and small size problems. There are many  
> options but I'm going to take one of the two options;
>
> Option1: ( 2 x Quadcore Intel Xeon 2.33 GHz CPUs, 16 giga memory) x  
> 4 computing nodes with a good server system + scsi hardware raid,  
> commercial compilers and softwares, Gigabit ethernet  = Total 32  
> CPU cores & 64 Giga RAM
>
> Option2: ( 2 x Quadcore Intel Xeon 2.00 GHz Cpus,  16 giga memory)  
> x 8 computing nodes with normal server system + SATA hardware/fake  
> raid, non-commercial compilers and softwares, Gigabit ethernet =  
> Total 64 CPU cores & 128 Giga RAM
>
> I prefer the second option but I don't know which one is better for  
> the purpose. The machine will be shared with several other members,  
> especailly for n-body simulations.
>
> I'd like to hear any of your experience and advice.

You could also by 6 nodes, allowing you to choose some intermediate  
options.

I suggest you run some tests to see whether the 2.33 GHz CPUs give  
you a performance improvement.  If so, that would be nice.  You  
should probably also test the non-commercial compilers and see  
whether they work well for your code, or whether there are problems.   
For example, many Linux distributions only ship older versions of  
gfortran, and you may need to build and test a newer version  
yourself.  You could also compare the speed of your code with  
different compilers.

Finally, for storage, it is very tempting to try and get away with a  
cheap solution.  I would definitely suggest a very reliable home  
partition for code and parameter files which is backed up daily.  A  
larger, less reliable scratch partition may work, but you will then  
have to live with the risk of losing data from time to time.  If this  
is a one-time purchase and there is no budget for repairs or  
replacements, you could consider buying something with an extended  
warranty, so that faulty parts are quickly replaced.

Personally, I prefer machines that are a bit smaller but more  
reliable instead.  Of course, this depends largely on whether you  
have hardware experts (e.g. interested and experienced students)  
nearby who can help you correct problems, or whether you have to do  
everything yourself.

-erik

-- 
Erik Schnetter <schnetter at cct.lsu.edu>

My email is as private as my paper mail.  I therefore support encrypting
and signing email messages.  Get my PGP key from www.keyserver.net.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 186 bytes
Desc: This is a digitally signed message part
Url : /archives/developers/attachments/20071115/8d8a5115/attachment.pgp