4

The scenario is as follows: Servers are equipped with two PSU's each. One of these is connected to the UPS, the other one directly to the power grid (bypassing the UPS).

I would assume that since the PSU's are redundant, load is more or less evenly distributed among these two PSU's 50/50 under normal conditions.

In the case of failure of the power grid (or UPS maintenance), one would think that the load now has to be beared by one PSU alone, amounting to an current load increase of 100% on this line (at least on the according bus).

Is that correct, or have you made different observations? Since I'm no electrical engineer, I don't know if perhaps manufacturers don't run 2 PSU's in full active-active mode but instead have some funky switching going on internally.

Roman
  • 3,907
  • 3
  • 21
  • 34

2 Answers2

7

Your understanding is correct: In modern servers power supplies are what is referred to as "Parallel Redundant" - Both power supplies are active, and both are supplying power to the system bus.

In the event of a power supply failure (internal fault, loss of input power, etc.) the remaining power supply (or power supplies) will need to supply additional current to the system bus to support the load (proportionally increasing their current draw from whatever their input source is - usually AC power).

You can see this behavior for yourself if you have a load-metering power strip (or a pair of in-line load meters, like Kill-A-Watt meters:

  • Connect both power supplies to separate metered sources
    (This can be the A and B legs of a CDU, or individual Kill-A-Watt Meters)
  • Power the system on and let it come up and stabilize. Note the load on each source (meter).
  • Unplug one of the power supplies.
    You will notice the input current to the other PSU increases by roughly what the one you just failed was drawing.

This has an important practical implication for power capacity planning that leads to a general rule of thumb: Never load a power circuit above 50% of its capacity.
This allows you to have the companion power leg fail and the load transfer over without risk of blowing a breaker/fuse.

(In reality you often don't want to load a power circuit above 40-45% of its capacity, as operating at exactly 100% capacity can be risky territory. Keeping the maximum potential single-fault load between 80% and 90% gives you some wiggle room.)

voretaq7
  • 79,879
  • 17
  • 130
  • 214
  • 1
    The NEC says that you shouldn't load up a circuit over 80% capacity. APC PDUs have an audible alarm if you hit this threshold. Also note that power supplies are not 100% efficient. Loading to 45% per circuit would likely be acceptable given that one PSU will get powered off in a failure. The second circuit would then likely be loaded to 80% due to fewer losses in the PSUs. – toppledwagon Feb 25 '13 at 18:09
  • @toppledwagon My experience with fail-overs is that when you have a fully populated rack drop a leg you gain a few tenths of an amp (mostly through meter rounding error) so I'd tend to err on the conservative side. Similarly my experience with electrical codes / load limits in (retail co-location) datacenters is that they're all too frequently ignored until a floor breaker pops during a maintenance window (at which point there's a massive re-balancing of power draw, and I get to say "I told you so!" :-P). – voretaq7 Feb 25 '13 at 18:16
  • It's interesting to hear your real world experiences on this. I would have thought the differences to be greater given the efficiency curves I've seen for the power supplies today. Maybe I'll get a chance to get some real data on this soon. Hopefully I'll be able to post our findings. – toppledwagon Feb 25 '13 at 21:15
  • @toppledwagon The CDUs we used at the colo have a 0.1 amp resolution on the local display, and take a few seconds to settle in on a final load figure. I think it's more meter error than PSU efficiency. We're not talking a huge amount of power either (0.3A on a 30A circuit). Other events (like power-cycling a box with a bunch of disks in it) could easily put you into overload though with startup current inrush (yet another reason to leave conservative safety margins) – voretaq7 Feb 25 '13 at 21:23
-2

Usually there is only 1 psu active.

Jeroen
  • 1,341
  • 7
  • 16