Well this depends on very much on your plans.
If you have hundred of boxes than you should run jobtracker and ndfs
name node not on the same machine, but I run it normally on one machine.
Also I run datanode and tasktrackers at the same time on the same
machine.
Reading the index from the ndfs is a bad idea since it is to slow,
but there was a discussion how to improve that and there will be a
solution soon.
I suggest to read this:
http://www.acmqueue.com/modules.php?name=Content&pa=showpage&pid=143
Stefan
Am 15.12.2005 um 14:31 schrieb Hamza Kaya:
It's obvious that high speed networks will be of great value. Are
there any
ideas on how to setup the system. I mean how should I distribute
the index?
Should ndfs servers be isolated from other jobs [e.g. JobTraker,
Distributed
Search Server or even fetching]?
Thanks.
--
Hamza
On 12/15/05, Stefan Groschupf <[EMAIL PROTECTED]> wrote:
... fast network 1.GBit and very low lantency is they most important
from my point of view.
Take many and cheap boxes (ata or sata hdds) but invest in your
network.
HTH
Stefan
Am 15.12.2005 um 13:51 schrieb Hamza Kaya:
Hi,
Currently I set up MapReduce on a 4 node cluster. Everything seems
working.
I was able to run Distributed Search also. However, I wonder
whether there
is a recommended cluster architecture set up that will boost the
performance. Or are there any previous work on cluster systems?
Thanks in advance.
--
Hamza Kaya
