[Beowulf] commercial clusters
hahn at physics.mcmaster.ca
Thu Sep 28 09:29:55 PDT 2006
> Is there many clients for processor time? As I saw the biggest
do you mean "how much real money will people pay for cycles
on clusters"? I don't know, but rumor has it that Sun's
$1/cpu-hour approach is not a resounding success. of course,
they're also not offering a real HPC cluster.
> supercomputers in the World are very busy! I'm wondering if it's
> worthwhile to setup a commercial cluster. Intel are planning for new
> processors - two CPUs each with quad cores. Two such machines will
> have power like one 50 GHz CPU:-)
Intel's got nothing special, at least not obviously over AMD.
it's also quite clear that multi-core has downsides - you certainly
can't sell multicore time as if each CPU has its full complement of
memory bandwidth, for instance.
if you consider $1/cpu-hour, that's $8766/year, which seems like
quite a lot (I think Sun is only offering opteron cores and gigabit-ish
networking.) if you figure the node depreciates to zero in 3 years,
that's $26K, which is a heck of a lot. the raw purchase price should
be around $1k (per core). where I live, electricity, incl cooling
would run ~$50/core for the whole period. I can't believe that suitable
space would make the overall cost > $2k.
so what we know is that at $1/cpu-hour (10-20x profit margin)
is too high to bring in hordes of customers,
at least given the services Sun offers.
it could be that tweaking the service would help (use virtualization,
probably, and provide some serious guarantees about performance),
but $1/cpu-hour is also probably a non-starter.
More information about the Beowulf