Dual CPU nodes?
math at velocet.ca
Tue Oct 22 15:53:15 EDT 2002
On Tue, Oct 22, 2002 at 11:52:12AM -0400, Robert G. Brown's all...
> On Tue, 22 Oct 2002, Ken Chase wrote:
> > On Mon, Oct 21, 2002 at 11:28:09AM -0400, Robert G. Brown's all...
> > > Most standard circuits have enough capacity to run two duals or four
> > > singles - you can estimate that they'll draw somewhere in the vicinity
> > > of 400-600 Watts, which is roughly a third the capacity of a twenty amp
> > > circuit for a switching power supply load, more like half the capacity
> > > of a 15 amp circuit. Then you've got a monitor and network switch,
> > > which can be another 100-200 Watts. A beefy UPS is not a bad thing to
> > wow. thats huge. I had 8 dual 2466s with 1.3tbirds on a 15Amp circuit (no
> > disk). When I moved it to a new building, I am guessing that the voltage was a
> > bit different, and the power bars they were wired into started blowing.
> > (perhaps they were drawing more amps at a lower voltage than my install
> > area?) At any rate, I put them in 8 per 20 amp circuit (instead of my
> > original intall 8/15) and nothing has blown (oh, and only 4 per power bar
> > too now :). Are you spec'ing for some seriously heavy disk power loads?
> My recent experiences with 2466's and building wiring are in the recent
> list archives and they are a sorry mess, although they are in the
> process of running the additional neutral wires required to get us
> perhaps semistable.
Agreed. Well we have a proper datacentre here and hired some electricians
with proper data centre wiring experience. So im reasonably confident
it was installed properly, but Im not the one around here who understands
how this is so and how to verify it independently. ;)
> We are finding that we can run at most 8 per 20 amp circuit here as
> well. One reason for the constriction in number of systems relative to
> their average power draw is that the switching power supplies draw a
> much higher PEAK current to provide the average power than one expects
> -- they have a relatively poor power factor. Apparently the circuit
> breakers used are sensitive to peak current more than rms current.
> Perhaps your 15 amp circuit used slow-blow breakers that were more
> sensitive to rms current than peak? The danger of doing this is that
> one can dangerously overload primary transformers and cause them to
> prematurely burn out. There are URL's in the stuff I posted last week
> (and some of the extended discussion that has occurred on this subject)
> that explain all of this.
Right. We are starting to look at better power factor power supplies, which
our regular parts supplier seems to have access to. They're about twice
the price for a .99PF supply than an Enermax at the same rating. Possibly
worth it over the years of operation and being nice to the transformers.
> I now just assume that any system I plug in has a power factor no better
> than about 0.8 (and will thus draw a peak current that is 25% or better
> high) and, because of the harmonic distortion of the supply voltage
> sinusoid, prefer to keep circuits relatively underloaded. We've seen a
> LOT of power related problems on circuits loaded close to capacity,
> although the non-shared neutrals may help this (I hope).
> > [A friend told me about blowing a circuit once at his lab and he had a large
> > metal bracelet on. When the circuit blew, he was standing under the power
> > run (about 5 feet over his head in the ceiling) and his bracelet shot up his
> > arm and yanked on his wrist, quite hard! He had a bruise to prove it. It
> > was only a 20A circuit - what would cause that? Dangerous wiring situation?]
> I have no idea. The current in the overhead line should have been
> balanced -- hot and neutral currents opposing in the conduit, the
> conduit itself grounded to boot to shield whatever field the resulting
> inductive surge would have caused. I don't see how an inductive EMP
> could have been generated strong enough to create a current in a
> bracelet that would magnetically attract it to the dropping line
Makes you wonder huh? well, the office power probably needs rewiring, the
data centre was obviously redone properly.
> Unless, perhaps, the overhead power run contained a single, unshielded
> high current line that was grounded at the panel, or there was a
> transformer nearby, and the pulse from this was responsible. Even then
> I'd be surprised that such a strong impulse force was generated -- 20
> amps isn't all that strong a current, so somewhere a much larger current
> must have been induced.
Exactly why I was asking the list. Seems far too large a current to drag
someone's arm in the air from 5 feet away.
> But I'm not a wiring expert. Maybe this is normal.
No, _SCARY_, not normal ;) He says the bracelet also got warm at the time ;)
> > > For that reason you might well want to select systems based on their low
> > > power requirements rather than their speed...
> > Has anyone here actually done this based on a full TCO analysis where
> > operations cost were as large (or larger) a factor than initial install?
> > Are TMTA's worth it in this situ?
> Sorry, TCO I get, but TMTA's?
sorry, the stock symbol for my poor stock - transmeta ;) I type that
so often on IRC I forgot.
Found this, however:
Im just wondering what the ratio of FLOPs to power is for these things...
of course, space is a factor too in every case, so you have to make
your own TCO spreadsheet if thats relevant to you. But a simple one
is power to computations executed. (And, as Feynman reminds us,
the more computations/s, the more heat you must generate, not a linear
relationship either! :) [ Then again, heat generated by the actual entropy
increase is tiny compared to the wasted heat of using clocks and pushing
electrons around - what are we still some 10+ orders of magnitude above
the minimum energy required to process a bit? -- Can you tell Ive
been reading "The Pleasure of Finding Things Out"? :) ]
> Robert G. Brown http://www.phy.duke.edu/~rgb/
> Duke University Dept. of Physics, Box 90305
> Durham, N.C. 27708-0305
> Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
Ken Chase, math at velocet.ca * Velocet Communications Inc. * Toronto, CANADA
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf