[Beowulf] harmonic mitigators vs PFC?

Robert G. Brown rgb at phy.duke.edu
Mon May 24 05:29:46 PDT 2004

On Sun, 23 May 2004, Mark Hahn wrote:

> Hi all.
> we're doing a new machineroom, which will dissipate around 250 KW.
> I'm having a hard time addressing the question of whether such a load
> (all modern PFC computers) needs harmonic mitigation.  the electricians
> all just say "your loads are nonlinear so you need HM".  I can't help
> think that the cost of HM hardware translates to more than a couple more nodes!
> and merely because the PSU is implemented with nonlinear components doesn't
> mean that the load is noisy.
> to my thinking, the active PFC found in current computers means precisely
> that HM is not necessary, since a PFC of .97 (so says my KillAWatt) indicates
> that only 3% of the power is drawn outside the ideal sine envelope.
> further, we're talking about O(800) seperate power supplies, and their 
> harmonics are probably not perfectly synchronized, and so sub-additive.
> can anyone offer advice or references on this?  we can apparently get a full
> answer by hiring a power consultant to bring in some kind of fancy power
> digitizer which will give us a plot of our load waveforms and presumably also
> a spectrum.  so far, we're going along with the HM plan, but mainly because
> it'll clean up the power coming in, and perhaps permit us to ride out some 
> flickers (the compute nodes won't have UPS-protected power...)

You're lucky your electricians are that clueful.  Ours weren't...

There is a HM link on the brahma resources page, let's see:


Their FAQ is one of the best explanations of HM and nonlinear loads
associated with switching power supplies that I've seen.

When last I looked at this (not too long ago) not so many power supplies
were PFC and the ones that were cost more (as one would expect).  The
argument advanced by Mirus is that putting a high quality HM transformer
at the room itself saves you this marginal cost (x N nodes) and keeps
you from having to make a PFC power supply a design specification for
future nodes.

Being a low suspicious sort of person, I'd also put a scope on the lines
and look at the waveforms for the voltage and current (ideally
simultaneously) before concluding that the current draw and voltage are
in sync and that there are no e.g. 180 Hz harmonics.  I'd guess that
this is what the fancy consultant is offering to sell you, but for what
you're likely to have to pay them you can probably buy an oscilloscope
and do it yourself.  I trust the kill-a-watt, sort of, but I trust my
eyes more.

You also should think of the whole picture (as it looks like you are
doing) and not just the PFC issue.  I believe that it is desirable to
have the server room transformer as local as possible and have the
distribution panel grounded to building steel (there are code rules on
just where and how things have to be grounded) to avoid ground loops and
excessively long runs to ground.  If you need a transformer anyway, then
you are really looking at the MARGINAL cost of the HM on the transformer
you get, and I don't think it is all that great.  HM can save you
operational money if you DO have non-PFC supplies -- the peak currents
drawn in the middle third of the phase a) cause the primary transformer
to run hot; b) heat is money; c) it can actually cause the primary
transformer to burn out prematurely which is a LOT of money; d) if you
run circuits at max capacity without correction, you typically brown out
the voltage during the peak due to resistive drops on the supply line,
which in turn causes the supply to run with the capacitor not properly
charged, which in turn reduces the system's fault tolerance.

Shoot your electrical contractor if they even THINK of sharing a common
neutral among three phases, and get them to use at least 12/2 throughout
even for "short" runs.  Thick wire stays cooler -- enough to pay for the
difference in cost several times over, I believe, over the ~10+ year
lifetime of such a space running 24x7 -- and minimizes the voltage on
the neutral at the receptacle relative to ground.  They'll likely bitch,
because bending thicker wire through conduits is somewhat more difficult
than thinner wire.  I think I read somewhere in my researches on power
that it is desirable in server room designs to keep the length of the
runs between distribution panel and receptacle down to 25 feet.

I don't know about the riding out flickers thing, but I would guess that
the HM transformers would do a pretty good job of buffering at least
certain classes of spikes and supply line wobbles.


> thanks, mark hahn.
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

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

More information about the Beowulf mailing list