At 20:12 09-06-98 -0400, Adam wrote:
>This is more of a curiosity question than anything. I got to thinking the
>other day, how do they figure out the Mean Time Before Failure of a
>harddrive. If you break out your calculator, most major manufactureres are
>30 years and upwards! For example, Western Digital lists a lot of their
>drives as 350,000 hourse(350,000 hours/24 hours in a day/365 days in a year
>= 39.95 years!!). I REALLY don't see this as at all possible for actual use.

Here's what IBM has to say about MTBF:

1. What is MTBF?

MTBF is the mean of a distribution of product life times, often estimated
by dividing the total operating time accumulated by a "defined group" of
drives within a given time period, by the total number of failures in that
time period.

2. What is this "defined group" of drives?

This is a group of drives that:

have not reached end-of-life (typically five to seven years)
are operated within a specified reliability temperature range, under
specified normal usage conditions, and
have not been damaged or abused.

3. What is considered to be a failure?

Any event that prevents a drive from performing its specified operations,
given that the drive meets the group definition described in question 2.

This includes drives that fail during shipment and during what is frequently
referred to as the "early life period" (failures typically resulting from
manufacturing defects).

It does not include drives that fail during integration into OEM system
units or as a result of mishandling, nor does it encompass drives that fail
beyond end-of-life.

4. If I purchase a drive with an MTBF of 1,000,000 hours (114
years), can I expect the drive will operate without failure for
1,000,000 hours?

No, because the drive will reach end-of-life before reaching 1,000,000
hours. For example, a continuously operated drive with a five-year useful
life will reach end-of-life in less than 45,000 hours. But, theoretically, if
the drive is replaced with a new drive when it reaches end-of-life, and the
new drive is replaced with another new drive when it reaches end-of-life,
etc., then the probability that 1,000,000 hours would elapse before a
failure occurs would be greater than 30 percent in most cases.

5. If I purchase 1000 drives with an MTBF of 1,000,000 hours,
how many can I expect to fail over a five-year period?

Assuming that any failed drive is replaced with a new drive having the
same reliability characteristics and that the drives are used continuously,
then the number of failures, r, you can expect is:
*
*                                                hours
*                      (1000 drives) x (43,800 ----------)
*                                                drive
*                r ~ ---------------------------------------------
*                                                hours
*                                  1,000,000 -------------
*                                                failure
*
*                  ~ 44 failures
*

Note that this number is subject to statistical variation.

In this example, because of statistical variation, there is approximately a
90 percent probability that the actual number of failures will be between
33 and 55.

If the drives are operated for 16 hours per day instead of 24 hours per
day, then the number of failures you can expect is:
*
*
*
*                                                hours
*                     (1000 drives) x (29,200 ----------)
*                                                drive
*                 r ~ ---------------------------------------------
*                                                hours
*                                 1,000,000 -------------
*                                                failure
*
*                   ~ 29 failures
*

6. IBM reports a "predicted MTBF." What does this mean?

It is very costly and time-consuming actually to measure high MTBFs
with a reasonable degree of precision. Therefore, to assess the reliability
of a new disk drive prior to volume production, reliability data from past
products and component and assembly tests is merged to create a
mathematical model of the drive reliability characteristics. The outcome of
the modeling process is the "predicted MTBF." After volume production
gets under way, actual field failure data is used to check the validity of the
model.

7. If I buy drives that have a "predicted MTBF" of 1,000,000
hours, can I expect to achieve 1,000,000 MTBF from those
drives?

Yes, given the conditions stated in question 2. The actual MTBF
measured from any specific set of drives will depend on the usage and the
environmental conditions the drives experience.

Stressing a drive beyond normal usage conditions may reduce the actual
MTBF to a point below the "predicted MTBF." Generally, reliability
decreases as temperature increases, so drives that are operated in warm
environments with poor airflow will tend to have a lower MTBF than
those operated in cool environments with good air flow. Drives that
experience a high seek rate tend to have a somewhat lower reliability than
those that experience a low seek rate. And drives that are in portable
equipment tend to be subject to higher levels of shock and vibration,
which also degrades reliability.

Furthermore, because MTBF can only be measured using statistical
methods, any measurement will be subject to statistical variation. The
degree of variation depends on the number of drives included in the
measurement. With more devices, less variation can be expected.

Regards,
Bill