[Beowulf] re: power prediction and planning
hahn at mcmaster.ca
Mon Oct 5 00:14:42 PDT 2009
in reference to recent mention of TDP and power planning,
what do you think the trend in the next few years will be?
5 years ago, people were talking about exponential increases
in power and power density - afaikt with straight faces.
that clearly didn't happen. in fact, dissipation has clearly
gone down (on a per-node basis - power per flop has gone through
I'm looking at current nodes and seeing substantial benefits to
improvements in cpu power, lower voltage/power ram, more sensible
chipsets, more efficient power supplies. we have a lot of nodes
bought ~4 years ago - dual-socket HP DL145G2's. cpu power was
around 100W/socket, and I suspect PSUs were probably ~80% efficient.
now it would be hard to miss 92-93% efficient PSUs and CPUs
in the 60-80W range. I'd be surprised if support chips haven't
improved as well. I know even disks are dissipating a lot less,
even though they never were a large component.
I've even seen some vendors brag about how their fans are more
efficient (which resonates with me because when a DL145G2 gets warm,
the ten (10) fans ramp up dissipate noticably more power,
which doesn't help the room temperature at all...)
I'm also not really talking about heat density - I generally assume
that we'll get an average of 2 nodes per 1U - ~80/rack.
Intel seems to be agressively pushing fab tech as well - .32 nm is
due this year, and they claim significant speed/power advances.
SO: do you expect pretty much constant per-node dissipation?
if higher, why? if lower, savings due to what?
we've planned with 300W/node for a long time, but current nodes are
noticably cooler than that. I imagine next-gen nodes will be lower
power, not higher, unless something else changes (4 socket becomes
common, or every node needs a Fermi or Larrabee card...)
also, does anyone have thoughts about machineroom features that
would support multiple generations, preferably without a lot of reno?
for instance, we have mostly L6-30 power, which seems reasonably safe
for commodity systems. but how about rack-back water cooling systems?
what are the chances of getting multiple generations out of that
(at least the heat-rejection part, if not the pipes or server-side
heat-exchangers.) is anyone working on commodity-ish systems
with water/coolant going to each node? (that is, skip or minimize
the use of air)
thanks, mark hahn.
More information about the Beowulf