[Beowulf] Advice on 4 CPU node configurations

[Cris Rhea]

> From: Prentice Bisbal <prentice at ias.edu>
> Subject: Re: [Beowulf] Advice on 4 CPU node configurations
> 
> Hearns, John wrote:
> >>   We are designing a new HPC system for our lab. We are shifting ideas
> >> around (our budget is of 30K euro).
> >>
> 
> 30K EUR won't get you much, especially if you have to by IB  or 10 GbE
> networking hardware. For that little money, I would recommend buying a
> 4-socket server with 8- or 12-core CPUs, and load it up with as much RAM
> as possible. You might be able to by 2 or 3, depending on how much you
> max them out.
> 
> If you go this route, make sure it has redundant disks, power supplies
> and all this, since you won't have inherent redunancy of duplicate
> cluster nodes.
> 
> I recently ordered a couple of Dell PowerEdge R815s with 32-cores (4
> sockets) and 128 GB of RAM, and my users have been flocking to them.
> Unfortunately, one of these servers keeps reporting SBEs since October
> and that still hasn't resolved to my satisfaction (documented elsewhere
> on this list), so I wouldn't recommend that particular model.
> 
> -- 
> Prentice


I recently purchased a few servers from Advanced HPC: Supermicro RM-206
servers with 4 AMD sockets (currently using 4 x 12-core at 2.3GHz)
and 512GB (yes, 0.5TB) memory.

Heck of a machine for about the same price we paid for a Dell 905 
a couple years ago.

While not your typical "cluster node", the 48 CPU cores and
huge memory are a big win for "piggy" applications. 

Obviously, you can bottleneck on the network connection or the local
disks, but if the app can fit on a single machine, rather than having
to be split across several, it can be a win.

--- Cris

-- 
 Cristopher J. Rhea
 Mayo Clinic - Research Computing Facility
 200 First St SW, Rochester, MN 55905
 crhea at Mayo.EDU
 (507) 284-0587