[GE users] Execution hosts: Blades vs Individual Servers

Neil Baker neil.baker at crl.toshiba.co.uk
Tue Jul 29 11:31:52 BST 2008


Hi GridEngine Users,

 

I was wondering if anyone could recommend a manufacturer and machine type
for our new grid's execution hosts?  

 

I'm used to using machines with two CPUs in them (Xeon 3GHz CPUs with
hyper-threading), running 4 jobs per machine at once, with 4GB RAM (1GB per
job) through a single Gigabit network card.  The data to be processed is all
stored on a very fast NAS device.

 

Of course I can't buy machines with these CPUs in them anymore.  When buying
a new machine with either dual core and quad core CPUs I have no idea how to
compare them against what we have.  I've heard that although they don't run
as fast as 3GHz, they are in fact more powerful for floating point
calculation, so would they process the jobs faster even though running
slower?  From what I've read and various stats I've seen, benchmarks show
how a Dual Core CPU is more powerful than a single core machine for running
one application (i.e. a game or a particular benchmark), but I can't seem to
find any data on how well they work at running one application (job) per CPU
(i.e. 2 jobs on a dual core, or 4 jobs on a quad core).

 

Intel always used to offer more resilient CPUs (we had some AMD chips burn
out 5 years ago), but are AMD just as reliable as Intel now?  As we're a
research department we have limited funding, so we're looking to get maximum
CPU power for our money, so if AMD are now reliable does anyone know if they
are still cheaper than the equivalent Intel CPU for the same performance?

 

Then we have Blades vs Indivudual Servers.  A Blade Server's back plane
seems like a single point of failure which could take out all the blades
within it, but do they offer a significant energy and heat saving to make
them a worthy option?

 

Sorry for all these questions and any help offered will be really
appreciated.

 

Kind Regards

 

Neil

 

 

 

 

 




More information about the gridengine-users mailing list