[Beowulf] Re: Estimating cluster power consumption - I/O issues

[Jim Lux]

At 01:40 PM 12/16/2005, Ed Karns wrote:
>" ... and being able to boot and run each system individually from
>its own hard drive will save you a world of time and possible
>frustration. ..."
>
>" ... ... Some may argue the exact opposite. ..."
>
>" ... ... ... Well, it *is* one of the most fundamental religious
>arguments of high performance clusters. ..."
>
>Granted and granted. It would seem that in the initial stages of
>building and bringing up a new beowulf system by a relatively new
>comer, the ability to boot from hard drive each individual node with
>potentially buggy code ... would seem to be an advantage.


If the nodes are highly heterogenous, true.  However, one big advantage of 
diskless is that you don't have the configuration management thing (let's 
see now, did I remember to push the latest version of my 
program/kernel/script out to all the nodes?)   The hassles I've had with 
bringing up small clusters (and there are many such hassles) have not been 
with getting diskless to work, rather the contrary, getting the same image 
on all the nodes. Once you figure out how to get PXE to do its thing, it's 
pretty bullet proof, and then, all your testing and development can be in 
one place, and you just push the reset button on the node to load the new 
version from the central repository.

As always though YMMV.




James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875