Re: Building a file server - advice please

Peter T. Breuer wrote (in part):

> And then he goes on to say exactly what I said - namely that the failure
> distributions may not be independent, and that the mtbf is only an
> expected value in a distribution, not a guarantee! (he doesn't say that
> there is high variability, so you get that from me for free).
>
It is funny to me how little people understand statistics, especially where
system reliability is concerned. In a former life, the company I worked for
received a request-for-proposal for a system specified by a U.S.Government
agency. In addition to all the rest of the specifications, was one that was
worded something like this:

The mean time between failures (MTBF) shall be 100,000 hours. This shall be
demonstrated by taking the first production unit an testing it. Unless it
operates for the specified amount of time, it will not meet these requirements.

I.e., even if we made 1000 units and they all ran for an 100 hours (which we
could never afford to do), this would not comply with the specification.

First of all, by the time the test was completed, the government agency
would have no use for the equipment. Second, perhaps the first unit would
fail at 1000 hours, and all the rest would go for 1,000,000. It would depend
on the distribution of the failures of such a system. But the guy who wrote
the specification clearly thought that the MTBF specification guaranteed
that all units would last at least that long, which is ignorant. No matter
what the MTBF, one (or a few) unit(s) could fail very early.

Consider ordinary incandescent light bulbs. Their MTTF (MTBF does not really
apply to light bulbs, because they are not repaired, but replaced) is about
750 hours, except for the little 7.5 watt night lights that go around 1200
hours, and "long life" bulbs that may go longer. But this is only if you
turn them on, leave them on, and control the input voltage better than my
power company does. And these, once the initial failures due to congenital
weakness are over, seem to follow a normal distribution. But if you turn
them on and off a lot, the mean comes down rapidly, and they seem to fail as
soon as you turn them on, when they get a large current surge.

--
.~. Jean-David Beyer Registered Linux User 85642.
/V\ PGP-Key: 9A2FC99A Registered Machine 241939.
/( )\ Shrewsbury, New Jersey http://counter.li.org
^^-^^ 09:40:00 up 11 days, 26 min, 3 users, load average: 4.14, 4.13, 4.09
.

Relevant Pages

  • Re: Weibull Trouble
    ... distribution after running a life test with no failures. ... It seems as if you want to do statistical Power analysis ... "problem" is that of running a life-test with no failures. ... tested with failures and should have a similar shape parameter. ...
    (sci.stat.math)
  • Re: Does anyone have a good definition of MTBF ?
    ... MTBF is the (Total number of unit operating hours) divided by the total ... It is important not to confuse MTBF with "life". ... for random failures, not wearout/burnout. ...
    (sci.electronics.misc)
  • Re: Weibull Trouble
    ... Basically I'm struggling to derive a scale parameter for aWeibull ... distribution after running a life test with no failures. ... "problem" is that of running a life-test with no failures. ... tested with failures and should have a similar shape parameter. ...
    (sci.stat.math)
  • Re: Humax Freeview HDD box UHF loopthrough
    ... But there are ageing phenomena in electronics which occur whether it's powered up or not. ... With very few exceptions, the ageing mechanisms are very slow indeed, which is impressive for something with such tiny manufacturing costs. ... you buy a piece of kit with a measured MTBF of 10 years, ... Such is the way of random failures. ...
    (uk.tech.digital-tv)
  • Re: Interpreting the test result on a RNG
    ... distribution as an approximation to the binomial distribution. ... the number of the tested sequences can be 50 or 100. ... CUMis the probability to see more than F failures. ... rejecting a perfect random-number generator for having ...
    (sci.crypt)