Appears in the Proceedings of the 5th USENIX Conference on File and Storage Technologies (FASTā?07), February 2007
Failure Trends in a Large Disk Drive Population
Eduardo Pinheiro, Wolf-Dietrich Weber and Luiz AndrĀ“ Barroso
e
Google Inc.
1600 Amphitheatre Pkwy
Mountain View, CA 94043
{edpin,wolf,luiz}@google.com
Abstract for guiding the design of storage systems as well as de-
vising deployment and maintenance strategies.
It is estimated that over 90% of all new information produced Despite the importance of the subject, there are very
in the world is being stored on magnetic media, most of it on few published studies on failure characteristics of disk
hard disk drives. Despite their importance, there is relatively
drives. Most of the available information comes from
little published work on the failure patterns of disk drives, and
the disk manufacturers themselves [2]. Their data are
the key factors that affect their lifetime. Most available data
typically based on extrapolation from accelerated life
are either based on extrapolation from accelerated aging exper-
test data of small populations or from returned unit
iments or from relatively modest sized ļ¬eld studies. Moreover,
databases. Accelerated life tests, although useful in pro-
larger population studies rarely have the infrastructure in place
viding insight into how some environmental factors can
to collect health signals from components in operation, which
affect disk drive lifetime, have been known to be poor
is critical information for detailed failure analysis.
predictors of actual failure rates as seen by customers
We present data collected from detailed observations of a
in the ļ¬eld [7]. Statistics from returned units are typi-
large disk drive population in a production Internet services de-
cally based on much larger populations, but since there
ployment. The population observed is many times larger than
that of previous studies. In addition to presenting failure statis- is little or no visibility into the deployment characteris-
tics, we analyze the correlation between failures and several tics, the analysis lacks valuable insight into what actu-
parameters generally believed to impact longevity. ally happened to the drive during operation. In addition,
Our analysis identiļ¬es several parameters from the driveās since units are typically returned during the warranty pe-
self monitoring facility (SMART) that correlate highly with riod (often three years or less), manufacturersā? databases
failures. Despite this high correlation, we conclude that mod- may not be as helpful for the study of long-term effects.
els based on SMART parameters alone are unlikely to be useful
A few recent studies have shed some light on ļ¬eld
for predicting individual drive failures. Surprisingly, we found
failure behavior of disk drives [6, 7, 9, 16, 17, 19, 20].
that temperature and activity levels were much less correlated
However, these studies have either reported on relatively
with drive failures than previously reported.
modest populations or did not monitor the disks closely
enough during deployment to provide insights into the
1 Introduction factors that might be associated with failures.
Disk drives are generally very reliable but they are
The tremendous advances in low-cost, high-capacity also very complex components. This combination
magnetic disk drives have been among the key factors means that although they fail rarely, when they do fail,
helping establish a modern society that is deeply reliant the possible causes of failure can be numerous. As a
on information technology. High-volume, consumer- result, detailed studies of very large populations are the
grade disk drives have become such a successful prod- only way to collect enough failure statistics to enable
uct that their deployments range from home computers meaningful conclusions. In this paper we present one
and appliances to large-scale server farms. In 2002, for such study by examining the population of hard drives
example, it was estimated that over 90% of all new in- under deployment within Googleās computing infras-
formation produced was stored on magnetic media, most tructure.
of it being hard disk drives [12]. It is therefore critical We have built an infrastructure that collects vital in-
to improve our understanding of how robust these com- formation about all Googleās systems every few min-
ponents are and what main factors are associated with utes, and a repository that stores these data in time-
failures. Such understanding can be particularly useful series format (essentially forever) for further analysis.
The information collected includes environmental fac-
tors (such as temperatures), activity levels and many of
the Self-Monitoring Analysis and Reporting Technology
(SMART) parameters that are believed to be good indi-
cators of disk drive health. We mine through these data
and attempt to ļ¬nd evidence that corroborates or con-
tradicts many of the commonly held beliefs about how
various factors can affect disk drive lifetime.
Our paper is unique in that it is based on data from a
disk population size that is typically only available from
vendor warranty databases, but has the depth of deploy-
ment visibility and detailed lifetime follow-up that only
an end-user study can provide. Our key ļ¬ndings are:
ā? Contrary to previously reported results, we found
very little correlation between failure rates and ei-
ther elevated temperature or activity levels.
ā? Some SMART parameters (scan errors, realloca-
tion counts, ofļ¬ine reallocation counts, and proba-
tional counts) have a large impact on failure proba-
bility.
Figure 1: Collection, storage, and analysis architecture.
ā? Given the lack of occurrence of predictive SMART
resources, error indications, and conļ¬guration informa-
signals on a large fraction of failed drives, it is un-
tion. It is imperative that this daemonās resource usage
likely that an accurate predictive failure model can
be very light, so not to interfere with the applications.
be built based on these signals alone.
One way to assure this is to have the machine-level col-
lector poll individual machines relatively infrequently
2 Background (every few minutes). Other slower changing data (such
as conļ¬guration information) and data from other exist-
ing databases can be collected even less frequently than
In this section we describe the infrastructure that was
that. Most notably for this study, data regarding ma-
used to gather and process the data used in this study,
chine repairs and disk swaps are pulled in from another
the types of disk drives included in the analysis, and in-
database.
formation on how they are deployed.
The System Health database is built upon Bigtable
[3], a distributed data repository widely used within
2.1 The System Health Infrastructure
Google, which itself is built upon the Google File Sys-
tem (GFS) [8]. Bigtable takes care of all the data layout,
The System Health infrastructure is a large distributed
compression, and access chores associated with a large
software system that collects and stores hundreds of
data store. It presents the abstraction of a 2-dimensional
attribute-value pairs from all of Googleās servers, and
table of data cells, with different versions over time mak-
provides the interface for arbitrary analysis jobs to pro-
ing up a third dimension. It is a natural ļ¬t for keeping
cess that data.
track of the values of different variables (columns) for
The architecture of the System Health infrastructure
different machines (rows) over time. The System Health
is shown in Figure 1. It consists of a data collection
database thus retains a complete time-ordered history of
layer, a distributed repository and an analysis frame-
the environment, utilization, error, conļ¬guration, and re-
work. The collection layer is responsible for getting in-
pair events in each machineās life.
formation from each of thousands of individual servers
Analysis programs run on top of the System Health
into a centralized repository. Different ļ¬avors of col-
database, looking at information from individual ma-
lectors exist to gather different types of data. Much of
chines, or mining the data across thousands of machines.
the health information is obtained from the machines di-
Large-scale analysis programs are typically built upon
rectly. A daemon runs on every machine and gathers
Googleās Mapreduce [5] framework. Mapreduce auto-
local data related to that machineās health, such as envi-
mates the mechanisms of large-scale distributed compu-
ronmental parameters, utilization information of various
tation (such as work distribution, load balancing, toler- the user site are found to have no defect by the manu-
ance of failures), allowing the user to focus simply on facturers upon returning the unit. Hughes et al. [11] ob-
the algorithms that make up the heart of the computa- serve between 20-30% āno problem foundā? cases after
tion. analyzing failed drives from their study of 3477 disks.
The analysis pipeline used for this study consists of From an end-userās perspective, a defective drive is
a Mapreduce job written in the Sawzall language and one that misbehaves in a serious or consistent enough
framework [15] to extract and clean up periodic SMART manner in the userās speciļ¬c deployment scenario that
data and repair data related to disks, followed by a pass it is no longer suitable for service. Since failures are
through R [1] for statistical analysis and ļ¬nal graph gen- sometimes the result of a combination of components
eration. (i.e., a particular drive with a particular controller or ca-
ble, etc), it is no surprise that a good number of drives
that fail for a given user could be still considered op-
2.2 Deployment Details erational in a different test harness. We have observed
that phenomenon ourselves, including situations where
The data in this study are collected from a large num-
a drive tester consistently āgreen lightsā? a unit that in-
ber of disk drives, deployed in several types of systems
variably fails in the ļ¬eld. Therefore, the most accurate
across all of Googleās services. More than one hundred
deļ¬nition we can present of a failure event for our study
thousand disk drives were used for all the results pre-
is: a drive is considered to have failed if it was replaced
sented here. The disks are a combination of serial and
as part of a repairs procedure. Note that this deļ¬nition
parallel ATA consumer-grade hard disk drives, ranging
implicitly excludes drives that were replaced due to an
in speed from 5400 to 7200 rpm, and in size from 80 to
upgrade.
400 GB. All units in this study were put into production
Since it is not always clear when exactly a drive failed,
in or after 2001. The population contains several models
we consider the time of failure to be when the drive was
from many of the largest disk drive manufacturers and
replaced, which can sometimes be a few days after the
from at least nine different models. The data used for
observed failure event. It is also important to mention
this study were collected between December 2005 and
that the parameters we use in this study were not in use
August 2006.
as part of the repairs diagnostics procedure at the time
As is common in server-class deployments, the disks
that these data were collected. Therefore there is no risk
were powered on, spinning, and generally in service for
of false (forced) correlations between these signals and
essentially all of their recorded life. They were deployed
repair outcomes.
in rack-mounted servers and housed in professionally-
managed datacenter facilities. Filtering. With such a large number of units monitored
Before being put into production, all disk drives go over a long period of time, data integrity issues invari-
through a short burn-in process, which consists of a ably show up. Information can be lost or corrupted along
combination of read/write stress tests designed to catch our collection pipeline. Therefore, some cleaning up of
many of the most common assembly, conļ¬guration, or the data is necessary. In the case of missing values, the
component-level problems. The data shown here do not individual values are marked as not available and that
include the fall-out from this phase, but instead begin speciļ¬c piece of data is excluded from the detailed stud-
when the systems are ofļ¬cially commissioned for use. ies. Other records for that same drive are not discarded.
Therefore our data should be consistent with what a reg- In cases where the data are clearly spurious, the entire
ular end-user should see, since most equipment manu- record for the drive is removed, under the assumption
facturers put their systems through similar tests before that one piece of spurious data draws into question other
shipment. ļ¬elds for the same drive. Identifying spurious data, how-
ever, is a tricky task. Because part of the goal of studying
the data is to learn what the numbers mean, we must be
2.3 Data Preparation
careful not to discard too much data that might appear
invalid. So we deļ¬ne spurious simply as negative counts
Deļ¬nition of Failure. Narrowly deļ¬ning what consti-
or data values that are clearly impossible. For exam-
tutes a failure is a difļ¬cult task in such a large opera-
ple, some drives have reported temperatures that were
tion. Manufacturers and end-users often see different
hotter than the surface of the sun. Others have had neg-
statistics when computing failures since they use differ-
ative power cycles. These were deemed spurious and
ent deļ¬nitions for it. While drive manufacturers often
removed. On the other hand, we have not ļ¬ltered any
quote yearly failure rates below 2% [2], user studies have
suspiciously large counts from the SMART signals, un-
seen rates as high as 6% [9]. Elerath and Shah [7] report
der the hypothesis that large counts, while improbable as
between 15-60% of drives considered to have failed at
raw numbers, are likely to be good indicators of some-
thing really bad with the drive. Filtering for spurious
values reduced the sample set size by less than 0.1%.
3 Results
We now analyze the failure behavior of our ļ¬eet of disk
drives using detailed monitoring data collected over a
nine-month observation window. During this time we
recorded failure events as well as all the available en-
vironmental and activity data and most of the SMART
parameters from the drives themselves. Failure informa-
tion spanning a much longer interval (approximately ļ¬ve
Figure 2: Annualized failure rates broken down by age groups
years) was also mined from an older repairs database.
All the results presented here were tested for their statis-
ulation. The higher baseline AFR for 3 and 4 year old
tical signiļ¬cance using the appropriate tests.
drives is more strongly inļ¬uenced by the underlying re-
liability of the particular models in that vintage than by
3.1 Baseline Failure Rates disk drive aging effects. It is interesting to note that our
3-month, 6-months and 1-year data points do seem to
Figure 2 presents the average Annualized Failure Rates
indicate a noticeable inļ¬uence of infant mortality phe-
(AFR) for all drives in our study, aged zero to 5 years,
nomena, with 1-year AFR dropping signiļ¬cantly from
and is derived from our older repairs database. The data
the AFR observed in the ļ¬rst three months.
are broken down by the age a drive was when it failed.
Note that this implies some overlap between the sample
sets for the 3-month, 6-month, and 1-year ages, because 3.2 Manufacturers, Models, and Vintages
a drive can reach its 3-month, 6-month and 1-year age
Failure rates are known to be highly correlated with drive
all within the observation period. Beyond 1-year there is
models, manufacturers and vintages [18]. Our results do
no more overlap.
not contradict this fact. For example, Figure 2 changes
While it may be tempting to read this graph as strictly
signiļ¬cantly when we normalize failure rates per each
failure rate with drive age, drive model factors are
drive model. Most age-related results are impacted by
strongly mixed into these data as well. We tend to source
drive vintages. However, in this paper, we do not show a
a particular drive model only for a limited time (as new,
breakdown of drives per manufacturer, model, or vintage
more cost-effective models are constantly being intro-
due to the proprietary nature of these data.
duced), so it is often the case that when we look at sets
Interestingly, this does not change our conclusions. In
of drives of different ages we are also looking at a very
contrast to age-related results, we note that all results
different mix of models. Consequently, these data are
shown in the rest of the paper are not affected signiļ¬?-
not directly useful in understanding the effects of disk
cantly by the population mix. None of our SMART data
age on failure rates (the exception being the ļ¬rst three
results change signiļ¬cantly when normalized by drive
data points, which are dominated by a relatively stable
model. The only exception is seek error rate, which is
mix of disk drive models). The graph is nevertheless a
dependent on one speciļ¬c drive manufacturer, as we dis-
good way to provide a baseline characterization of fail-
cuss in section 3.5.5.
ures across our population. It is also useful for later
studies in the paper, where we can judge how consistent
the impact of a given parameter is across these diverse
3.3 Utilization
drive model groups. A consistent and noticeable impact
across all groups indicates strongly that the signal being The literature generally refers to utilization metrics by
measured has a fundamentally powerful correlation with employing the term duty cycle which unfortunately has
failures, given that it is observed across widely varying no consistent and precise deļ¬nition, but can be roughly
ages and models. characterized as the fraction of time a drive is active out
The observed range of AFRs (see Figure 2) varies of the total powered-on time. What is widely reported in
from 1.7%, for drives that were in their ļ¬rst year of op- the literature is that higher duty cycles affect disk drives
eration, to over 8.6%, observed in the 3-year old pop- negatively [4, 21].
It is difļ¬cult for us to arrive at a meaningful numer-
ical utilization metric given that our measurements do
not provide enough detail to derive what 100% utiliza-
tion might be for any given disk model. We choose in-
stead to measure utilization in terms of weekly averages
of read/write bandwidth per drive. We categorize utiliza-
tion in three levels: low, medium and high, correspond-
ing respectively to the lowest 25th percentile, 50-75th
percentiles and top 75th percentile. This categorization
is performed for each drive model, since the maximum
bandwidths have signiļ¬cant variability across drive fam-
ilies. We note that using number of I/O operations and
bytes transferred as utilization metrics provide very sim-
ilar results. Figure 3 shows the impact of utilization on
AFR across the different age groups. Figure 3: Utilization AFR
Overall, we expected to notice a very strong and con-
sistent correlation between high utilization and higher 3.4 Temperature
failure rates. However our results appear to paint a more
complex picture. First, only very young and very old Temperature is often quoted as the most important envi-
age groups appear to show the expected behavior. Af- ronmental factor affecting disk drive reliability. Previous
ter the ļ¬rst year, the AFR of high utilization drives is studies have indicated that temperature deltas as low as
at most moderately higher than that of low utilization 15C can nearly double disk drive failure rates [4]. Here
drives. The three-year group in fact appears to have the we take temperature readings from the SMART records
opposite of the expected behavior, with low utilization every few minutes during the entire 9-month window
drives having slightly higher failure rates than high uti- of observation and try to understand the correlation be-
lization ones. tween temperature levels and failure rates.
One possible explanation for this behavior is the sur- We have aggregated temperature readings in several
vival of the ļ¬ttest theory. It is possible that the fail- different ways, including averages, maxima, fraction of
ure modes that are associated with higher utilization are time spent above a given temperature value, number of
more prominent early in the driveās lifetime. If that is the times a temperature threshold is crossed, and last tem-
case, the drives that survive the infant mortality phase perature before failure. Here we report data on averages
are the least susceptible to that failure mode, and result and note that other aggregation forms have shown sim-
in a population that is more robust with respect to varia- ilar trends and and therefore suggest the same conclu-
tions in utilization levels. sions.
Another possible explanation is that previous obser- We ļ¬rst look at the correlation between average tem-
vations of high correlation between utilization and fail- perature during the observation period and failure. Fig-
ures has been based on extrapolations from manufactur- ure 4 shows the distribution of drives with average tem-
ersā? accelerated life experiments. Those experiments are perature in increments of one degree and the correspond-
likely to better model early life failure characteristics, ing annualized failure rates. The ļ¬gure shows that fail-
and as such they agree with the trend we observe for the ures do not increase when the average temperature in-
young age groups. It is possible, however, that longer creases. In fact, there is a clear trend showing that lower
term population studies could uncover a less pronounced temperatures are associated with higher failure rates.
effect later in a driveās lifetime. Only at very high temperatures is there a slight reversal
of this trend.
When we look at these results across individual mod-
els we again see a complex pattern, with varying pat- Figure 5 looks at the average temperatures for differ-
terns of failure behavior across the three utilization lev- ent age groups. The distributions are in sync with Figure
els. Taken as a whole, our data indicate a much weaker 4 showing a mostly ļ¬at failure rate at mid-range temper-
correlation between utilization levels and failures than atures and a modest increase at the low end of the tem-
previous work has suggested. perature distribution. What stands out are the 3 and 4-
year old drives, where the trend for higher failures with
higher temperature is much more constant and also more
pronounced.
Overall our experiments can conļ¬rm previously re-
ones. At the end of this section we discuss our results
and reason about the usefulness of SMART parameters
in obtaining predictive models for individual disk drive
failures.
We present results in three forms. First we compare
the AFR of drives with zero and non-zero counts for a
given parameter, broken down by the same age groups
as in ļ¬gures 2 and 3. We also ļ¬nd it useful to plot the
probability of survival of drives over the nine-month ob-
servation window for different ranges of parameter val-
ues. Finally, in addition to the graphs, we devise a sin-
gle metric that could relay how relevant the values of
a given SMART parameter are in predicting imminent
failures. To that end, for each SMART parameter we
Figure 4: Distribution of average temperatures and failures look for thresholds that increased the probability of fail-
rates. ure in the next 60 days by at least a factor of 10 with
respect to drives that have zero counts for that parame-
ter. We report such Critical Thresholds whenever we are
able to ļ¬nd them with high conļ¬dence (> 95%).
3.5.1 Scan Errors
Drives typically scan the disk surface in the background
and report errors as they discover them. Large scan error
counts can be indicative of surface defects, and therefore
are believed to be indicative of lower reliability. In our
population, fewer than 2% of the drives show scan errors
and they are nearly uniformly spread across various disk
models.
Figure 6 shows the AFR values of two groups of
drives, those without scan errors and those with one or
Figure 5: AFR for average drive temperature. more. We plot bars across all age groups in which we
have statistically signiļ¬cant data. We ļ¬nd that the group
ported temperature effects only for the high end of our of drives with scan errors are ten times more likely to fail
temperature range and especially for older drives. In the than the group with no errors. This effect is also noticed
lower and middle temperature ranges, higher tempera- when we further break down the groups by disk model.
tures are not associated with higher failure rates. This is From Figure 8 we see a drastic and quick decrease in
a fairly surprising result, which could indicate that data- survival probability after the ļ¬rst scan error (left graph).
center or server designers have more freedom than pre- A little over 70% of the drives survive the ļ¬rst 8 months
viously thought when setting operating temperatures for after their ļ¬rst scan error. The dashed lines represent the
equipment that contains disk drives. We can conclude 95% conļ¬dence interval. The middle plot in Figure 8
that at moderate temperature ranges it is likely that there separates the population in four age groups (in months),
are other effects which affect failure rates much more and shows an effect that is not visible in the AFR plots. It
strongly than temperatures do. appears that scan errors affect the survival probability of
young drives more dramatically very soon after the ļ¬rst
scan error occurs, but after the ļ¬rst month the curve ļ¬at-
3.5 SMART Data Analysis
tens out. Older drives, however, continue to see a steady
decline in survival probability throughout the 8-month
We now look at the various self-monitoring signals that
period. This behavior could be another manifestation of
are available from virtually all of our disk drives through
infant mortality phenomenon. The right graph in ļ¬gure 8
the SMART standard interface. Our analysis indicates
looks at the effect of multiple scan errors. While drives
that some signals appear to be more relevant to the study
with one error are more likely to fail than those with
of failures than others. We ļ¬rst look at those in detail,
none, drives with multiple errors fail even more quickly.
and then list a summary of our ļ¬ndings for the remaining
Figure 6: AFR for scan errors. Figure 7: AFR for reallocation counts.
Figure 8: Impact of scan errors on survival probability. Left ļ¬gure shows aggregate survival probability for all drives after ļ¬rst
scan error. Middle ļ¬gure breaks down survival probability per drive ages in months. Right ļ¬gure breaks down drives by their
number of scan errors.
The critical threshold analysis conļ¬rms what the groups (Figure 7), even if slightly less pronounced.
charts visually imply: the critical threshold for scan er- Drives with one or more reallocations do fail more of-
rors is one. After the ļ¬rst scan error, drives are 39 times ten than those with none. The average impact on AFR
more likely to fail within 60 days than drives without appears to be between a factor of 3-6x.
scan errors. Figure 11 shows the survival probability after the ļ¬rst
reallocation. We truncate the graph to 8.5 months, due
to a drastic decrease in the conļ¬dence levels after that
3.5.2 Reallocation Counts
point. In general, the left graph shows, about 85% of the
When the driveās logic believes that a sector is damaged drives survive past 8 months after the ļ¬rst reallocation.
(typically as a result of recurring soft errors or a hard er- The effect is more pronounced (middle graph) for drives
ror) it can remap the faulty sector number to a new phys- in the age ranges [10,20) and [20, 60] months, while
ical sector drawn from a pool of spares. Reallocation newer drives in the range [0,5) months suffer more than
counts reļ¬ect the number of times this has happened, their next generation. This could again be due to infant
and is seen as an indication of drive surface wear. About mortality effects, although it appears to be less drastic in
9% of our population has reallocation counts greater this case than for scan errors.
than zero. Although some of our drive models show After their ļ¬rst reallocation, drives are over 14 times
higher absolute values than others, the trends we observe more likely to fail within 60 days than drives without
are similar across all models. reallocation counts, making the critical threshold for this
As with scan errors, the presence of reallocations parameter also one.
seems to have a consistent impact on AFR for all age
Figure 9: AFR for ofļ¬ine reallocation count. Figure 10: AFR for probational count.
Figure 11: Impact of reallocation count values on survival probability. Left ļ¬gure shows aggregate survival probability for all
drives after ļ¬rst reallocation. Middle ļ¬gure breaks down survival probability per drive ages in months. Right ļ¬gure breaks down
drives by their number of reallocations.
3.5.3 Ofļ¬ine Reallocations points were not within high conļ¬dence intervals). Drives
in the older age groups appear to be more highly affected
Ofļ¬ine reallocations are deļ¬ned as a subset of the real- by it, although we are unable to attribute this effect to
location counts studied previously, in which only real- age given the different model mixes in the various age
located sectors found during background scrubbing are groups.
counted. In other words, it should exclude sectors that
After the ļ¬rst ofļ¬ine reallocation, drives have over
are reallocated as a result of errors found during actual
21 times higher chances of failure within 60 days than
I/O operations. Although this deļ¬nition mostly holds,
drives without ofļ¬ine reallocations; an effect that is
we see evidence that certain disk models do not imple-
again more drastic than total reallocations.
ment this deļ¬nition. For instance, some models show
Our data suggest that, although ofļ¬ine reallocations
more ofļ¬ine reallocations than total reallocations. Since
could be an important parameter affecting failures, it is
the impact of ofļ¬ine reallocations appears to be signif-
particularly important to interpret trends in these values
icant and not identical to that of total reallocations, we
within speciļ¬c models, since there is some evidence that
decided to present it separately (Figure 9). About 4% of
different drive models may classify reallocations differ-
our population shows non-zero values for ofļ¬ine reallo-
ently.
cations, and they tend to be concentrated on a particular
subset of drive models.
3.5.4 Probational Counts
Overall, the effects on survival probability of ofļ¬ine
reallocation seem to be more drastic than those of to-
Disk drives put suspect bad sectors āon probationā? un-
tal reallocations, as seen in Figure 12 (as before, some
til they either fail permanently and are reallocated or
curves are clipped at 8 months because our data for those
continue to work without problems. Probational counts,
Figure 12: Impact of ofļ¬ine reallocation on survival probability. Left ļ¬gure shows aggregate survival probability for all drives
after ļ¬rst ofļ¬ine reallocation. Middle ļ¬gure breaks down survival probability per drive ages in months. Right ļ¬gure breaks down
drives by their number ofļ¬ine reallocation.
Figure 13: Impact of probational count values on survival probability. Left ļ¬gure shows aggregate survival probability for all
drives after ļ¬rst probational count. Middle ļ¬gure breaks down survival probability per drive ages in months. Right ļ¬gure breaks
down drives by their number of probational counts.
therefore, can be seen as a softer error indication. It 3.5.5 Miscellaneous Signals
could provide earlier warning of possible problems but
In addition to the SMART parameters described in the
might also be a weaker signal, in that sectors on pro-
previous sections, which we have found to most closely
bation may indeed never be reallocated. About 2% of
impact failure rates, we have also studied several other
our drives had non-zero probational count values. We
parameters from the SMART set as well as other envi-
note that this number is lower than both online and of-
ronmental factors. Here we brieļ¬y mention our relevant
ļ¬ine reallocation counts, likely indicating that sectors
ļ¬ndings for some of those parameters.
may be removed from probation after further observa-
tion of their behavior. Once more, the distribution of Seek Errors. Seek errors occur when a disk drive fails to
drives with non-zero probational counts are somewhat properly track a sector and needs to wait for another rev-
skewed towards a subset of disk drive models. olution to read or write from or to a sector. Drives report
Figures 10 and 13 show that probational count trends it as a rate, and it is meant to be used in combination with
are generally similar to those observed for ofļ¬ine re- model-speciļ¬c thresholds. When examining our popu-
allocations, with age group being somewhat less pro- lation, we ļ¬nd that seek errors are widespread within
nounced. The critical threshold for probational counts drives of one manufacturer only, while others are more
is also one: after the ļ¬rst event, drives are 16 times more conservative in showing this kind of errors. For this one
likely to fail within 60 days than drives with zero proba- manufacturer, the trend in seek errors is not clear, chang-
tional counts. ing from one vintage to another. For other manufactur-
ers, there is no correlation between failure rates and seek
errors.
CRC Errors. Cyclic redundancy check (CRC) errors
are detected during data transmission between the phys- 3.5.6 Predictive Power of SMART Parameters
ical media and the interface. Although we do observe
Given how strongly correlated some SMART parame-
some correlation between higher CRC counts and fail-
ters were found to be with higher failure rates, we were
ures, those effects are somewhat less pronounced. CRC
hopeful that accurate predictive failure models based on
errors are less indicative of drive failures than that of ca-
SMART signals could be created. Predictive models are
bles and connectors. About 2% of our population had
very useful in that they can reduce service disruption
CRC errors.
due to failed components and allow for more efļ¬cient
Power Cycles. The power cycles indicator counts the scheduled maintenance processes to replace the less ef-
number of times a drive is powered up and down. In ļ¬cient (and reactive) repairs procedures. In fact, one of
a server-class deployment, in which drives are powered the main motivations for SMART was to provide enough
continuously, we do not expect to reach high enough insight into disk drive behavior to enable such models to
power cycle counts to see any effects on failure rates. be built.
Our results ļ¬nd that for drives aged up to two years, this After our initial attempts to derive such models
is true, there is no signiļ¬cant correlation between fail- yielded relatively unimpressive results, we turned to the
ures and high power cycles count. But for drives 3 years question of what might be the upper bound of the accu-
and older, higher power cycle counts can increase the racy of any model based solely on SMART parameters.
absolute failure rate by over 2%. We believe this is due Our results are surprising, if not somewhat disappoint-
more to our population mix than to aging effects. More- ing. Out of all failed drives, over 56% of them have no
over, this correlation could be the effect (not the cause) count in any of the four strong SMART signals, namely
of troubled machines that require many repair iterations scan errors, reallocation count, ofļ¬ine reallocation, and
and thus many power cycles to be ļ¬xed. probational count. In other words, models based only
on those signals can never predict more than half of the
Calibration Retries. We were unable to reach a consis-
failed drives. Figure 14 shows that even when we add
tent and clear deļ¬nition of this SMART parameter from
all remaining SMART parameters (except temperature)
public documents as well as consultations with some of
we still ļ¬nd that over 36% of all failed drives had zero
the disk manufacturers. Nevertheless, our observations
counts on all variables. This population includes seek
do not indicate that this is a particularly useful parame-
error rates, which we have observed to be widespread in
ter for the goals of this study. Under 0.3% of our drives
our population (> 72% of our drives have it) which fur-
have calibration retries, and of that group only about 2%
ther reduces the sample size of drives without any errors.
have failed, making this a very weak and imprecise sig-
nal when compared with other SMART parameters. It is difļ¬cult to add temperature to this analysis since
despite it being reported as part of SMART there are no
Spin Retries. Counts the number of retries when the
crisp thresholds that directly indicate errors. However,
drive is attempting to spin up. We did not register a sin-
if we arbitrarily assume that spending more than 50%
gle count within our entire population.
of the observed time above 40C is an indication of pos-
sible problem, and add those drives to the set of pre-
Power-on hours Although we do not dispute that
dictable failures, we still are left with about 36% of all
power-on hours might have an effect on drive lifetime,
drives with no failure signals at all. Actual useful mod-
it happens that in our deployment the age of the drive is
els, which need to have small false-positive rates are in
an excellent approximation for that parameter, given that
fact likely to do much worse than these limits might sug-
our drives remain powered on for most of their life time.
gest.
Vibration This is not a parameter that is part of the
We conclude that it is unlikely that SMART data alone
SMART set, but it is one that is of general concern in de-
can be effectively used to build models that predict fail-
signing drive enclosures as most manufacturers describe
ures of individual drives. SMART parameters still ap-
how vibration can affect both performance and reliabil-
pear to be useful in reasoning about the aggregate reli-
ity of disk drives. Unfortunately we do not have sensor
ability of large disk populations, which is still very im-
information to measure this effect directly for drives in
portant for logistics and supply-chain planning. It is pos-
service. We attempted to indirectly infer vibration ef-
sible, however, that models that use parameters beyond
fects by considering the differences in failure rates be-
those provided by SMART could achieve signiļ¬cantly
tween systems with a single drive and those with mul-
better accuracies. For example, performance anomalies
tiple drives, but those experiments were not controlled
and other application or operating system signals could
enough for other possible factors to allow us to reach
be useful in conjunction with SMART data to create
any conclusions.
more powerful models. We plan to explore this possi-
bility in our future work.
ployments than a typical disk drive vendor might have.
Although they do not report directly on the correlation
between SMART parameters or environmental factors
and failures (possibly for conļ¬dentiality concerns), their
work is useful in enabling a qualitative understanding
of factors what affect disk drive reliability. For exam-
ple, they comment that end-user failure rates can be as
much as ten times higher than what the drive manufac-
turer might expect [7]; they report in [6] a strong experi-
mental correlation between number of heads and higher
failure rates (an effect that is also predicted by the mod-
els in [4]); and they observe that different failure mech-
anisms are at play at different phases of a drive lifetime
[19]. Generally, our ļ¬ndings are in line with these re-
sults.
Figure 14: Percentage of failed drives with SMART errors. User experience studies may lack the depth of insight
into the device inner workings that is possible in man-
4 Related Work ufacturer reports, but they are essential in understand-
ing device behavior in real-world deployments. Unfortu-
nately, there are very few such studies to date, probably
Previous studies in this area generally fall into two cat-
due to the large number of devices needed to observe
egories: vendor (disk drive or storage appliance) tech-
statistically signiļ¬cant results and the complex infras-
nical papers and user experience studies. Disk ven-
tructure required to track failures and their contributing
dors studies provide valuable insight into the electro-
factors.
mechanical characteristics of disks and both model-
based and experimental data that suggests how several Talagala and Patterson [20] perform a detailed er-
environmental factors and usage activities can affect de- ror analysis of 368 SCSI disk drives over an eighteen
vice lifetime. Yang and Sun [21] and Cole [4] describe month period, reporting a failure rate of 1.9%. Re-
the processes and experimental setup used by Quantum sults on a larger number of desktop-class ATA drives
and Seagate to test new units and the models that attempt under deployment at the Internet Archive are presented
to make long-term reliability predictions based on accel- by Schwarz et al [17]. They report on a 2% failure rate
erated life tests of small populations. Power-on-hours, for a population of 2489 disks during 2005, while men-
duty cycle, temperature are identiļ¬ed as the key deploy- tioning that replacement rates have been as high as 6%
ment parameters that impact failure rates, each of them in the past. Gray and van Ingen [9] cite observed fail-
having the potential to double failure rates when going ure rates ranging from 3.3-6% in two large web prop-
from nominal to extreme values. For example, Cole erties with 22,400 and 15,805 disks respectively. A re-
presents thermal de-rating models showing that MTBF cent study by Schroeder and Gibson [16] helps shed light
could degrade by as much as 50% when going from op- into the statistical properties of disk drive failures. The
erating temperatures of 30C to 40C. Coleās report also study uses failure data from several large scale deploy-
presents yearly failure rates from Seagateās warranty ments, including a large number of SATA drives. They
database, indicating a linear decrease in annual failure report a signiļ¬cant overestimation of mean time to fail-
rates from 1.2% in the ļ¬rst year to 0.39% in the third ure by manufacturers and a lack of infant mortality ef-
(and last year of record). In our study, we did not ļ¬nd fects. None of these user studies have attempted to cor-
much correlation between failure rate and either elevated relate failures with SMART parameters or other environ-
temperature or utilization. It is the most surprising result mental factors.
of our study. Our annualized failure rates were generally We are aware of two groups that have attempted
higher than those reported by vendors, and more consis- to correlate SMART parameters with failure statistics.
tent with other user experience studies. Hughes et al [11, 13, 14] and Hamerly and Elkan [10].
Shah and Elerath have written several papers based The largest populations studied by these groups was of
on the behavior of disk drives inside Network Appli- 3744 and 1934 drives and they derive failure models that
ance storage products [6, 7, 19]. They use a reliability achieve predictive rates as high as 30%, at false posi-
database that includes ļ¬eld failure statistics as well as tive rates of about 0.2% (that false-positive rate corre-
support logs, and their position as an appliance vendor sponded to a number of drives between 20-43% of the
enables them more control and visibility into actual de- drives that actually failed in their studies). Hughes et al.
Acknowledgments
also cites an annualized failure rate of 4-6%, based on
their 2-3 month long experiment which appears to use
stress test logs provided by a disk manufacturer. We wish to acknowledge the contribution of numer-
ous Google colleagues, particularly in the Platforms and
Our study takes a next step towards a better under-
Hardware Operations teams, who made this study pos-
standing of disk drive failure characteristics by essen-
sible, directly or indirectly; among them: Xiaobo Fan,
tially combining some of the best characteristics of stud-
Greg Slaughter, Don Yang, Jeremy Kubica, Jim Winget,
ies from vendor database analysis, namely population
Caio Villela, Justin Moore, Henry Green, Taliver Heath,
size, with the kind of visibility into a real-world deploy-
and Walt Drummond. We are also thankful to our shep-
ment that is only possible with end-user data.
herd, Mary Baker for comments and guidance. A special
thanks to Urs HĀØ lzle for his extensive feedback on our
o
5 Conclusions drafts.
In this study we report on the failure characteristics of
References
consumer-grade disk drives. To our knowledge, the
study is unprecedented in that it uses a much larger
[1] The r project for statistical computing.
population size than has been previously reported and
http://www.r-project.org.
presents a comprehensive analysis of the correlation be-
tween failures and several parameters that are believed to
[2] Dave Anderson, Jim Dykes, and Erik Riedel. More
affect disk lifetime. Such analysis is made possible by
than an interface - scsi vs. ata. In Proceedings of
a new highly parallel health data collection and analysis
the 2nd USENIX Conference on File and Storage
infrastructure, and by the sheer size of our computing
Technologies (FASTā?03), pages 245 ā? 257, Febru-
deployment.
ary 2003.
One of our key ļ¬ndings has been the lack of a con-
sistent pattern of higher failure rates for higher temper- [3] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wil-
ature drives or for those drives at higher utilization lev- son C. Hsieh, Deborah A. Wallach, Mike Burrows,
els. Such correlations have been repeatedly highlighted Tushar Chandra, Andrew Fikes, and Robert E.
by previous studies, but we are unable to conļ¬rm them Gruber. Bigtable: A distributed storage system for
by observing our population. Although our data do not structured data. In Proceedings of the 7th USENIX
allow us to conclude that there is no such correlation, Symposium on Operating Systems Design and Im-
it provides strong evidence to suggest that other effects plementation (OSDIā?06), November 2006.
may be more prominent in affecting disk drive reliabil-
[4] Gerry Cole. Estimating drive reliability in desktop
ity in the context of a professionally managed data center
computers and consumer electronics systems. Sea-
deployment.
gate Technology Paper TP-338.1, November 2000.
Our results conļ¬rm the ļ¬ndings of previous smaller
population studies that suggest that some of the SMART [5] Jeffrey Dean and Sanjay Ghemawat. Mapre-
parameters are well-correlated with higher failure prob- duce: Simpliļ¬ed data processing on large clus-
abilities. We ļ¬nd, for example, that after their ļ¬rst scan ters. In Proceedings of the 6th USENIX Symposium
error, drives are 39 times more likely to fail within 60 on Operating Systems Design and Implementation
days than drives with no such errors. First errors in re- (OSDIā?04), pages 137 ā? 150, December 2004.
allocations, ofļ¬ine reallocations, and probational counts
[6] Jon G. Elerath and Sandeep Shah. Disk drive re-
are also strongly correlated to higher failure probabil-
liability case study: Dependence upon ļ¬y-height
ities. Despite those strong correlations, we ļ¬nd that
and quantity of heads. In Proceedings of the An-
failure prediction models based on SMART parameters
nual Symposium on Reliability and Maintainabil-
alone are likely to be severely limited in their prediction
ity, January 2003.
accuracy, given that a large fraction of our failed drives
have shown no SMART error signals whatsoever. This
[7] Jon G. Elerath and Sandeep Shah. Server class
result suggests that SMART models are more useful in
disk drives: How reliable are they? In Proceed-
predicting trends for large aggregate populations than for
ings of the Annual Symposium on Reliability and
individual components. It also suggests that powerful
Maintainability, pages 151 ā? 156, January 2004.
predictive models need to make use of signals beyond
those provided by SMART. [8] Sanjay Ghemawat, Howard Gobioff, and Shun-Tak
Leung. The google ļ¬le system. In Proceedings of
the 19th ACM Symposium on Operating Systems [18] Sandeep Shah and Jon G. Elerath. Disk drive vin-
Principles, pages 29 ā? 43, December 2003. tage and its effect on reliability. In Proceedings
of the Annual Symposium on Reliability and Main-
[9] Jim Gray and Catherine van Ingen. Empirical tainability, pages 163 ā? 167, January 2004.
measurements of disk failure rates and error rates.
Technical Report MSR-TR-2005-166, December [19] Sandeep Shah and Jon G. Elerath. Reliability anal-
2005. ysis of disk drive failure mechanisms. In Proceed-
ings of the Annual Symposium on Reliability and
[10] Greg Hamerly and Charles Elkan. Bayesian ap- Maintainability, pages 226 ā? 231, January 2005.
proaches to failure prediction for disk drives. In
Proceedings of the Eighteenth International Con- [20] Nisha Talagala and David Patterson. An analysis
ference on Machine Learning (ICMLā?01), June of error behavior in a large storage system. Techni-
2001. cal Report CSD-99-1042, University of California,
Berkeley, February 1999.
[11] Gordon F. Hughes, Joseph F. Murray, Kenneth
Kreutz-Delgado, and Charles Elkan. Improved [21] Jimmy Yang and Feng-Bin Sun. A comprehensive
disk-drive failure warnings. IEEE Transactions on review of hard-disk drive reliability. In Proceed-
Reliability, 51(3):350 ā? 357, September 2002. ings of the Annual Symposium on Reliability and
Maintainability, pages 403 ā? 409, January 1999.
[12] Peter Lyman and Hal R.Varian.
How much information? October
2003. http://www2.sims.berkeley.edu/
research/projects/how-much-info-2003/index.htm.
[13] Joseph F. Murray, Gordon F Hughes, and Kenneth
Kreutz-Delgado. Hard drive failure prediction us-
ing non-parametric statistical methods. Proceed-
ings of ICANN/ICONIP, June 2003.
[14] Joseph F. Murray, Gordon F. Hughes, and Ken-
neth Kreutz-Delgado. Machine learning methods
for predicting failures in hard drives: A multiple-
instance application. J. Mach. Learn. Res., 6:783ā?
816, 2005.
[15] Rob Pike, Sean Dorward, Robert Griesemer, and
Sean Quinlan. Interpreting the data: Parallel anal-
ysis with sawzall. Scientiļ¬c Programming Jour-
nal, Special Issue on Grids and Worldwide Com-
puting Programming Models and Infrastructure,
13(4):227 ā? 298.
[16] Bianca Schroeder and Garth A. Gibson. Disk
failures in the real world: What does an mttf of
1,000,000 hours mean to you? In Proceedings of
the 5th USENIX Conference on File and Storage
Technologies (FAST), February 2007.
[17] Thomas Schwartz, Mary Baker, Steven Bassi,
Bruce Baumgart, Wayne Flagg, Catherine van
Ingen, Kobus Joste, Mark Manasse, and Mehul
Shah. Disk failure investigations at the internet
archive. 14th NASA Goddard, 23rd IEEE Confer-
ence on Mass Storage Systems and Technologies,
May 2006.
|