[gmx-users] strategy of clustering 10nodes ?

[Anton Feenstra]

Taeho Kim wrote:
> Hi
> We'd like to set up our own cluster, none commercial products, firstly 10 nodes and
> go large later.  If you are in a situation with 2users and 10nodes cluster, how could
> you split jobs / nodes?   i.e. clustering nodes for multi-jobs like 2jobs (4nodes-6nodes,
> 8-2), 3jobs (4-4-2), or 4jobs(3-3-2-2) ?  Is there anyone have similar system/situation? 

It depends on a number of factors. The main issue is about scaling, which
depends on the type and size of system you simulate. You will need to try
this a few times with different types of systems you are going to run.
In general, you would take as many CPU's per job that still give good
scaling, say >70% or >80%. That number of CPU's then determines how many
of these jobs you could run at the same time.

You can also approach it from the other end, and take the number of simulations
that you (and your coleagues) want to perform. If this is equal to or larger
than the number of CPU's you have, you could run each simulation on a single
CPU. Each job will run longer, but all together will be faster than when
running each of them on multiple CPU's.


-- 
Groetjes,

Anton
 _____________ _______________________________________________________
|             |                                                       |
|  _   _  ___,| K. Anton Feenstra                                     |
| / \ / \'| | | Dept. of Pharmacochem. - Vrije Universiteit Amsterdam |
|(   |   )| | | De Boelelaan 1083 - 1081 HV Amsterdam - Netherlands   |
| \_/ \_/ | | | Tel: +31 20 44 47608 - Fax: +31 20 44 47610           |
|             | Feenstra at chem.vu.nl - www.chem.vu.nl/~feenstra/       |
|             | "If You See Me Getting High, Knock Me Down"           |
|             | (Red Hot Chili Peppers)                               |
|_____________|_______________________________________________________|