[Beowulf] Opteron

[laytonjb at charter.net]

---- Joe Landman <landman at scalableinformatics.com> wrote: 
> Mark Hahn wrote:
> >> Come on then SC07'ers. Whats the buzz with barcelona?
> > 

> 
> I am not sure about this.  Have some tests to do.  Bug Doug Eadline on 
> this, he has some good early impressions based upon his testing.

The initial results are amazing! I'm sure Doug will chime but he's on an
airplane right now (probably asleep or showing off his EeePC to the entire
airplane).

Doug has shared some of the results with me (NAS Parallel Benchmarks).
I've also seen some Barcelona 1.9 results for the same basic configuration.
While I don't want to steal Doug's thunder... the 2.33 GHz Hapertown
just killed the 2.66 GHz Clovertown. It's the same motherboard, same
memory, same FSB, etc. The only change was the CPU.

The Barcelona results were very disappointing. I played some
spreadsheet games and scaled the Barcelona and Harpertown results
to 3.0 GHz, scaling with clock speed. If the basic trneds hold true, the
Barcelona was only able to beat Harpertown on 1 of the NPB's. But, let
me also say, that I don't know how these benchmarks scale with clock
speed, so I'm sure my results are a bit off-base. Plus, I've heard that the
PGI compiler can really help the Barcelona (Doug can tell us all about the
details of his testing, but he was very conservative with his Harpertown
results, which makes them even more amazing).

> 
> > any benefit from the ATI merger is more than a few years off, IMO,
> 
> Some of us doubt there ever will be any benefit from this merger.
> 
> > and quite specialized.  Fusion may be a useful feature for markets where 
> > a highly integrated solution is rewarded (cheap desktops, laptops).
> > so far, at least, it's just "chip legos", which is not what I'd call
> > a Big Vision.  Intel's reported approach of x86-ISA-based GPUs sounds
> > MUCH more exciting, since it could be hopefully be subverted to do HPC
> > (and without the sadistic computing model of current GP-GPU.)

Actually, I don't have too much trouble with the basic concepts. Translating
them into actual code is a bit more of a challenge. CUDA really, really helps
in this regard. Plus I know a couple of the Nvidia developers who write code
using CUDA and they tell me that it's dead easy (these are friends, not just
SC07 friends, so when they tell me something I really believe it).

> > I'm more interested in hearing whether 10G eth made a good showing
> > vs IB.
> 
> I saw the Mellanox 10GbE units, Myricom's stuff,  the Arastra (sp?) 
> switches, Woven systems stuff.  For 10GbE to be meaningful, it has to be 
> port cost competitive with IB.  Three of the vendors above grasped that. 
>   There were a number of vendors that (mistakenly) believe that $2k 
> USD/port is a good thing in clusters and storage systems.  I wish them 
> luck with that.  Myricom's stuff is pretty neat, well priced.  They have 
> some interesting ideas.

Ditto.

The 10GigE companies are still under the delusion that $2,000/port is acceptable.
I hate to tell them that for an interconnect, that's not true. I think the NICs are
OK in price (still a little high), but the switch prices are a mess. Way to expensive
to ever have a meaningful impact on a computational interconnect. I've been
hearing for the last 3 years that "this year is the year of 10GigE." I haven't
seen it yet. But then again, who knows? Maybe 2008 will be the year of
10GigE?

I think the more likely scenario is multi-fabric 10GigE and IB. These switches
are here and from what I've heard they work quite well. So if you need
10GigE, for example for storage, you can just plug it into a multi-fabric
switch. This is a much more likely situation IMHO.

Jeff