Mulling over MTBF.

[Donald Becker]

On Wed, 12 Sep 2001, Carlos O'Donell Jr. wrote:

> On a recent trend, I have been discussing with various
> colleagues the aspect of power usage and power saving
> in clusters.

I'm guessing that you have already looked at the Scyld features for
soft-power-down and Wake-On-LAN (WOL) wake-up.  Our system has the
advantage that we identify nodes by the station address, and thus already
have the information needed for WOL.

> At first, power saving, through node sleeping or drive
> spin down, seemed like a good idea.
> 
> Though, I am wary about the following effects:
> 
> - Does spindown/spinup on common IDE drives effect MTBF?

Yes.  Typical disk drive ratings put the spin-up count equivalent to
about 9 hours of the MTBF.  Those numbers are not directly comparable,
but it's a useful number to look at.

Laptop drives are typically set to spin down after a few minutes of idle
time, both because power savings are much more important and because a
stopped drive is more resistant to shocks.

> - Does node sleeping/wakeup cycles effect MTBF for voltage
>   supplies on the motherboard? (Or other componenets, through
>   relaxation and transients).

Not obviously: the HV side of most ATX power supplies is continuously
powered, so there is no inrush current shock coming out of stand-by
mode.  The thermal stress from the varying load is likely the dominant
effect.

We have one batch of ATX power supplies that are very likely to fail in
the brown-out conditions around a power failure.  Those same supplies
have not failed when the machine is in stand-by mode during power
failures.  Yes, the real solution is to get different power supplies,
however this is an example of soft-power-off increasing the MTBF of the
system. 

Donald Becker				becker at scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
Annapolis MD 21403			410-990-9993