I've been similarly surprised by how efficient modern boxes are. That said,
boxes with a large number of CPUs can have a wide range of power
requirements. We have some 64-core boxes that idle at ~150W but hit ~800W
under high CPU load. I've found HPL to be a useful tool to get CPUs
stressed out, but we're an HPC shop.

Skylar


On Thu, Jan 16, 2014 at 6:16 AM, Edward Ned Harvey (lopser) <
[email protected]> wrote:

>  In every server room, I always keep a spreadsheet showing how many VA
> and W each UPS is rated for, and how many VA and W each system draws, and
> which UPS's each server is connected to.  It's a pain because I don't
> blindly add up all the max power supply ratings (which are typically about
> 5x higher than the *actual* max power requirements of the box)...  In order
> to get the measurement, I have to plug a power strip into a kill-a-watt,
> and then plug both of the redundant power supplies of the box into the
> power strip, then stress the system for a couple of minutes, and record the
> measurements.  This effectively solves a few problems (a) under sizing
> UPS's, which I've seen bite other admins, whose data centers crash under
> load etc, and (b) over sizing UPS's, which is what most admins do, and
> wastes money and space, and creates unnecessary waste.
>
>
>
> Anyway, all that is just a tangent.  Here's what I really came to say:
>
>
>
> Yesterday, I turned on a new server, took my power measurements, and was
> blown away to find that the system drew a maximum of 110W and 120VA,
> including 10% margin for error.  Dang!  They're getting good at power
> reduction.  Not long ago, I expected less powerful servers to draw 300-500
> or so, *actual* measured peak.
>
>
>
> BTW, assuming a typical server with nothing but cpu, ram, and disks in it,
> the way I stress the system is to launch a zillion threads of "cat
> /dev/urandom > /dev/null" and gzip, until "top" shows me I'm CPU starved...
> I have found experimentally that stressing the disks and/or network and/or
> other IO makes no noticeable difference to power consumption.  If there
> were any other power hungry components such as GPU, I would need to find a
> way to stress that component as well.
>
> _______________________________________________
> Tech mailing list
> [email protected]
> https://lists.lopsa.org/cgi-bin/mailman/listinfo/tech
> This list provided by the League of Professional System Administrators
>  http://lopsa.org/
>
>
_______________________________________________
Tech mailing list
[email protected]
https://lists.lopsa.org/cgi-bin/mailman/listinfo/tech
This list provided by the League of Professional System Administrators
 http://lopsa.org/

Reply via email to