[Beowulf] recommendations for cluster upgrades

[Mark Hahn]

>>> I'm currently shopping around for a cluster-expansion and was shopping
>>> for options. Anybody out here who's bought new hardware in the  recent
>>> past? Any suggestions? Any horror stories?
>
> MH> I think the answer depends on time-frame.  if you're bw-sensitive,
> MH> you go nehalem for the time being.
>
> I'm really surprised that everyone just screams "Nehalem" - of course the
> platorm is the fastest that you can buy for money at the moment. It is
> the youngest design so that is not suprising.

it's not just age, but rather that in one step, Intel destroys the best
argument for buying AMD (for the past 6 years!)  instead of delivering 
half the bandwidth-per-core as AMD, they now deliver twice.

> But clustering is always about price/performance. And AMD doesn't look so
>bad there. Shanghai-pricing was lowered in February and another pricedrop is
>expected when Istanbul is launched.

really?  vendors always drop the price of aging chips, but does this really
effect system prices much/enough?

amusingly the AMD pricing page (http://www.amd.com/pricing) is currently
404, so maybe they _are_ preparing to run a sale ;)

>So Nehalem is faster but more
>expensive. For some Applications it is "more expensive" than it is
>"faster".

do you have numbers?

> Memory, CPUs and Chipsets for Intel are expensive - what you can buy from
>AMD is "old" hardware that is in the market now for quite some time. It is
>simply cheaper.

I don't follow.  old is cheap, no matter who makes it.  you seem to be 
suggesting that intel and amd hardware have a different trajectory;
I haven't noticed this myself.

> MH> for the near future, I'm interested to see how AMD's next-gen "istanbul" works out.
>
> It seems to be very interesting.

what do you mean?  I'm hoping it'll include core tweaks, modern memory 
support and perhaps even some improvement for 2s systems from snooping.
the increased core count is, if anything, regressive if the memory 
interface remains 2x pc2/6400 per socket (since gflop/gbyte balance goes down.)

> HT-Assist	- some kinde of snoop filter for Istanbul. Without this
>Istanbul probably wouldn't scale to 6cores at all. But real world
>performance seem to be interesting... so HT-Assist seem to make a
>difference.

do you have numbers?  also, why would snooping be relevant for 6-core anywa?
the coherency protocol within a socket is afaikt, totally unchanged by 
snooping (which is totally an L3/uncore thing, no?)