Appears in the Proceedings of the 5th USENIX Conference on File and Storage Technologies (FAST’07), February 2007 Failure Trends in a Large Disk Drive Population Eduardo Pinheiro, Wolf-Dietrich Weber and Luiz Andr´e Barroso Google Inc. 1600 Amphitheatre Pkwy Mountain View, CA 94043 {edpin,wolf,luiz}@google.com Abstract It is estimated that over 90% of all new information produced in the world is being stored on magnetic media, most of it on hard disk drives. Despite their importance, there is relatively little published work on the failure patterns of disk drives, and the key factors that affect their lifetime. Most available data are either based on extrapolation from accelerated aging experiments or from relatively modest sized field studies. Moreover, larger population studies rarely have the infrastructure in place to collect health signals from components in operation, which is critical information for detailed failure analysis. We present data collected from detailed observations of a large disk drive population in a production Internet services deployment. The population observed is many times larger than that of previous studies. In addition to presenting failure statistics, we analyze the correlation between failures and several parameters generally believed to impact longevity. Our analysis identifies several parameters from the drive’s self monitoring facility (SMART) that correlate highly with failures. Despite this high correlation, we conclude that models based on SMART parameters alone are unlikely to be useful for predicting individual drive failures. Surprisingly, we found that temperature and activity levels were much less correlated with drive failures than previously reported. 1 Introduction The tremendous advances in low-cost, high-capacity magnetic disk drives have been among the key factors helping establish a modern society that is deeply reliant on information technology. High-volume, consumergrade disk drives have become such a successful product that their deployments range from home computers and appliances to large-scale server farms. In 2002, for example, it was estimated that over 90% of all new information produced was stored on magnetic media, most of it being hard disk drives [12]. It is therefore critical to improve our understanding of how robust these components are and what main factors are associated with failures. Such understanding can be particularly useful for guiding the design of storage systems as well as devising deployment and maintenance strategies. Despite the importance of the subject, there are very few published studies on failure characteristics of disk drives. Most of the available information comes from the disk manufacturers themselves [2]. Their data are typically based on extrapolation from accelerated life test data of small populations or from returned unit databases. Accelerated life tests, although useful in providing insight into how some environmental factors can affect disk drive lifetime, have been known to be poor predictors of actual failure rates as seen by customers in the field [7]. Statistics from returned units are typically based on much larger populations, but since there is little or no visibility into the deployment characteristics, the analysis lacks valuable insight into what actually happened to the drive during operation. In addition, since units are typically returned during the warranty period (often three years or less), manufacturers’ databases may not be as helpful for the study of long-term effects. A few recent studies have shed some light on field failure behavior of disk drives [6, 7, 9, 16, 17, 19, 20]. However, these studies have either reported on relatively modest populations or did not monitor the disks closely enough during deployment to provide insights into the factors that might be associated with failures. Disk drives are generally very reliable but they are also very complex components. This combination means that although they fail rarely, when they do fail, the possible causes of failure can be numerous. As a result, detailed studies of very large populations are the only way to collect enough failure statistics to enable meaningful conclusions. In this paper we present one such study by examining the population of hard drives under deployment within Google’s computing infras- tructure. We have built an infrastructure that collects vital information about all Google’s systems every few minutes, and a repository that stores these data in timeseries format (essentially forever) for further analysis. The information collected includes environmental factors (such as temperatures), activity levels and many of the Self-Monitoring Analysis and Reporting Technology (SMART) parameters that are believed to be good indicators of disk drive health. We mine through these data and attempt to find evidence that corroborates or contradicts 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 deployment visibility and detailed lifetime follow-up that only an end-user study can provide. Our key findings are: • Contrary to previously reported results, we found very little correlation between failure rates and either elevated temperature or activity levels. • Some SMART parameters (scan errors, reallocation counts, offline reallocation counts, and probational counts) have a large impact on failure proba- bility. • Given the lack of occurrence of predictive SMART signals on a large fraction of failed drives, it is unlikely that an accurate predictive failure model can be built based on these signals alone. 2 Background In this section we describe the infrastructure that was used to gather and process the data used in this study, the types of disk drives included in the analysis, and information on how they are deployed. 2.1 The System Health Infrastructure The System Health infrastructure is a large distributed software system that collects and stores hundreds of attribute-value pairs from all of Google’s servers, and provides the interface for arbitrary analysis jobs to process that data. The architecture of the System Health infrastructure is shown in Figure 1. It consists of a data collection layer, a distributed repository and an analysis framework. The collection layer is responsible for getting information from each of thousands of individual servers into a centralized repository. Different flavors of collectors exist to gather different types of data. Much of the health information is obtained from the machines directly. A daemon runs on every machine and gathers local data related to that machine’s health, such as environmental parameters, utilization information of various Figure 1: Collection, storage, and analysis architecture. resources, error indications, and configuration information. It is imperative that this daemon’s resource usage be very light, so not to interfere with the applications. One way to assure this is to have the machine-level collector poll individual machines relatively infrequently (every few minutes). Other slower changing data (such as configuration information) and data from other existing databases can be collected even less frequently than that. Most notably for this study, data regarding machine repairs and disk swaps are pulled in from another database. The System Health database is built upon Bigtable [3], a distributed data repository widely used within Google, which itself is built upon the Google File System (GFS) [8]. Bigtable takes care of all the data layout, compression, and access chores associated with a large data store. It presents the abstraction of a 2-dimensional table of data cells, with different versions over time making up a third dimension. It is a natural fit for keeping track of the values of different variables (columns) for different machines (rows) over time. The System Health database thus retains a complete time-ordered history of the environment, utilization, error, configuration, and repair events in each machine’s life. Analysis programs run on top of the System Health database, looking at information from individual machines, or mining the data across thousands of machines. Large-scale analysis programs are typically built upon Google’s Mapreduce [5] framework. Mapreduce automates the mechanisms of large-scale distributed compu- tation (such as work distribution, load balancing, tolerance of failures), allowing the user to focus simply on the algorithms that make up the heart of the computa- tion. The analysis pipeline used for this study consists of a Mapreduce job written in the Sawzall language and framework [15] to extract and clean up periodic SMART data and repair data related to disks, followed by a pass through R [1] for statistical analysis and final graph gen- eration. 2.2 Deployment Details The data in this study are collected from a large number of disk drives, deployed in several types of systems across all of Google’s services. More than one hundred thousand disk drives were used for all the results presented here. The disks are a combination of serial and parallel ATA consumer-grade hard disk drives, ranging in speed from 5400 to 7200 rpm, and in size from 80 to 400 GB. All units in this study were put into production in or after 2001. The population contains several models from many of the largest disk drive manufacturers and from at least nine different models. The data used for this study were collected between December 2005 and August 2006. As is common in server-class deployments, the disks were powered on, spinning, and generally in service for essentially all of their recorded life. They were deployed in rack-mounted servers and housed in professionallymanaged datacenter facilities. Before being put into production, all disk drives go through a short burn-in process, which consists of a combination of read/write stress tests designed to catch many of the most common assembly, configuration, or component-level problems. The data shown here do not include the fall-out from this phase, but instead begin when the systems are officially commissioned for use. Therefore our data should be consistent with what a regular end-user should see, since most equipment manufacturers put their systems through similar tests before shipment. 2.3 Data Preparation Definition of Failure. Narrowly defining what constitutes a failure is a difficult task in such a large operation. Manufacturers and end-users often see different statistics when computing failures since they use different definitions for it. While drive manufacturers often quote yearly failure rates below 2% [2], user studies have seen rates as high as 6% [9]. Elerath and Shah [7] report between 15-60% of drives considered to have failed at the user site are found to have no defect by the manufacturers upon returning the unit. Hughes et al. [11] observe between 20-30% “no problem found” cases after analyzing failed drives from their study of 3477 disks. From an end-user’s perspective, a defective drive is one that misbehaves in a serious or consistent enough manner in the user’s specific deployment scenario that it is no longer suitable for service. Since failures are sometimes the result of a combination of components (i.e., a particular drive with a particular controller or cable, etc), it is no surprise that a good number of drives that fail for a given user could be still considered operational in a different test harness. We have observed that phenomenon ourselves, including situations where a drive tester consistently “green lights” a unit that invariably fails in the field. Therefore, the most accurate definition we can present of a failure event for our study is: a drive is considered to have failed if it was replaced as part of a repairs procedure. Note that this definition implicitly excludes drives that were replaced due to an upgrade. Since it is not always clear when exactly a drive failed, we consider the time of failure to be when the drive was replaced, which can sometimes be a few days after the observed failure event. It is also important to mention that the parameters we use in this study were not in use as part of the repairs diagnostics procedure at the time that these data were collected. Therefore there is no risk of false (forced) correlations between these signals and repair outcomes. Filtering. With such a large number of units monitored over a long period of time, data integrity issues invariably show up. Information can be lost or corrupted along our collection pipeline. Therefore, some cleaning up of the data is necessary. In the case of missing values, the individual values are marked as not available and that specific piece of data is excluded from the detailed studies. Other records for that same drive are not discarded. In cases where the data are clearly spurious, the entire record for the drive is removed, under the assumption that one piece of spurious data draws into question other fields for the same drive. Identifying spurious data, however, is a tricky task. Because part of the goal of studying the data is to learn what the numbers mean, we must be careful not to discard too much data that might appear invalid. So we define spurious simply as negative counts or data values that are clearly impossible. For example, some drives have reported temperatures that were hotter than the surface of the sun. Others have had negative power cycles. These were deemed spurious and removed. On the other hand, we have not filtered any suspiciously large counts from the SMART signals, under the hypothesis that large counts, while improbable as raw numbers, are likely to be good indicators of something 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 fleet 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 environmental and activity data and most of the SMART parameters from the drives themselves. Failure information spanning a much longer interval (approximately five years) was also mined from an older repairs database. All the results presented here were tested for their statistical significance using the appropriate tests. 3.1 Baseline Failure Rates Figure 2 presents the average Annualized Failure Rates (AFR) for all drives in our study, aged zero to 5 years, and is derived from our older repairs database. The data 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 a drive can reach its 3-month, 6-month and 1-year age all within the observation period. Beyond 1-year there is no more overlap. While it may be tempting to read this graph as strictly failure rate with drive age, drive model factors are strongly mixed into these data as well. We tend to source a particular drive model only for a limited time (as new, more cost-effective models are constantly being introduced), so it is often the case that when we look at sets of drives of different ages we are also looking at a very different mix of models. Consequently, these data are not directly useful in understanding the effects of disk age on failure rates (the exception being the first three data points, which are dominated by a relatively stable mix of disk drive models). The graph is nevertheless a good way to provide a baseline characterization of failures 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 drive model groups. A consistent and noticeable impact across all groups indicates strongly that the signal being measured has a fundamentally powerful correlation with failures, given that it is observed across widely varying ages and models. The observed range of AFRs (see Figure 2) varies from 1.7%, for drives that were in their first year of operation, to over 8.6%, observed in the 3-year old popFigure 2: Annualized failure rates broken down by age groups ulation. The higher baseline AFR for 3 and 4 year old drives is more strongly influenced by the underlying reliability of the particular models in that vintage than by disk drive aging effects. It is interesting to note that our 3-month, 6-months and 1-year data points do seem to indicate a noticeable influence of infant mortality phenomena, with 1-year AFR dropping significantly from the AFR observed in the first three months. 3.2 Manufacturers, Models, and Vintages Failure rates are known to be highly correlated with drive models, manufacturers and vintages [18]. Our results do not contradict this fact. For example, Figure 2 changes significantly when we normalize failure rates per each drive model. Most age-related results are impacted by drive vintages. However, in this paper, we do not show a breakdown of drives per manufacturer, model, or vintage due to the proprietary nature of these data. Interestingly, this does not change our conclusions. In contrast to age-related results, we note that all results shown in the rest of the paper are not affected significantly by the population mix. None of our SMART data results change significantly when normalized by drive model. The only exception is seek error rate, which is dependent on one specific drive manufacturer, as we discuss in section 3.5.5. 3.3 Utilization The literature generally refers to utilization metrics by employing the term duty cycle which unfortunately has no consistent and precise definition, but can be roughly characterized as the fraction of time a drive is active out of the total powered-on time. What is widely reported in the literature is that higher duty cycles affect disk drives negatively [4, 21]. It is difficult for us to arrive at a meaningful numerical utilization metric given that our measurements do not provide enough detail to derive what 100% utilization might be for any given disk model. We choose instead to measure utilization in terms of weekly averages of read/write bandwidth per drive. We categorize utilization in three levels: low, medium and high, corresponding 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 significant variability across drive families. We note that using number of I/O operations and bytes transferred as utilization metrics provide very similar results. Figure 3 shows the impact of utilization on AFR across the different age groups. Overall, we expected to notice a very strong and consistent correlation between high utilization and higher failure rates. However our results appear to paint a more complex picture. First, only very young and very old age groups appear to show the expected behavior. After the first year, the AFR of high utilization drives is at most moderately higher than that of low utilization drives. The three-year group in fact appears to have the opposite of the expected behavior, with low utilization drives having slightly higher failure rates than high utilization ones. One possible explanation for this behavior is the survival of the fittest theory. It is possible that the failure modes that are associated with higher utilization are more prominent early in the drive’s lifetime. If that is the case, the drives that survive the infant mortality phase are the least susceptible to that failure mode, and result in a population that is more robust with respect to variations in utilization levels. Another possible explanation is that previous observations of high correlation between utilization and failures has been based on extrapolations from manufacturers’ accelerated life experiments. Those experiments are likely to better model early life failure characteristics, and as such they agree with the trend we observe for the young age groups. It is possible, however, that longer term population studies could uncover a less pronounced effect later in a drive’s lifetime. When we look at these results across individual models we again see a complex pattern, with varying patterns of failure behavior across the three utilization levels. Taken as a whole, our data indicate a much weaker correlation between utilization levels and failures than previous work has suggested. Figure 3: Utilization AFR 3.4 Temperature Temperature is often quoted as the most important environmental factor affecting disk drive reliability. Previous studies have indicated that temperature deltas as low as 15C can nearly double disk drive failure rates [4]. Here we take temperature readings from the SMART records every few minutes during the entire 9-month window of observation and try to understand the correlation between temperature levels and failure rates. We have aggregated temperature readings in several different ways, including averages, maxima, fraction of time spent above a given temperature value, number of times a temperature threshold is crossed, and last temperature before failure. Here we report data on averages and note that other aggregation forms have shown similar trends and and therefore suggest the same conclu- sions. We first look at the correlation between average temperature during the observation period and failure. Figure 4 shows the distribution of drives with average temperature in increments of one degree and the corresponding annualized failure rates. The figure shows that failures do not increase when the average temperature increases. In fact, there is a clear trend showing that lower temperatures are associated with higher failure rates. Only at very high temperatures is there a slight reversal of this trend. Figure 5 looks at the average temperatures for different age groups. The distributions are in sync with Figure 4 showing a mostly flat failure rate at mid-range temperatures and a modest increase at the low end of the temperature distribution. What stands out are the 3 and 4year old drives, where the trend for higher failures with higher temperature is much more constant and also more pronounced. Overall our experiments can confirm previously re- Figure 4: Distribution of average temperatures and failures rates. Figure 5: AFR for average drive temperature. ported temperature effects only for the high end of our temperature range and especially for older drives. In the lower and middle temperature ranges, higher temperatures are not associated with higher failure rates. This is a fairly surprising result, which could indicate that datacenter or server designers have more freedom than previously thought when setting operating temperatures for equipment that contains disk drives. We can conclude that at moderate temperature ranges it is likely that there are other effects which affect failure rates much more strongly than temperatures do. 3.5 SMART Data Analysis We now look at the various self-monitoring signals that are available from virtually all of our disk drives through the SMART standard interface. Our analysis indicates that some signals appear to be more relevant to the study of failures than others. We first look at those in detail, and then list a summary of our findings for the remaining 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 figures 2 and 3. We also find it useful to plot the probability of survival of drives over the nine-month observation window for different ranges of parameter values. Finally, in addition to the graphs, we devise a single 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 look for thresholds that increased the probability of failure in the next 60 days by at least a factor of 10 with respect to drives that have zero counts for that parameter. We report such Critical Thresholds whenever we are able to find them with high confidence (> 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 more. We plot bars across all age groups in which we have statistically significant data. We find that the group of drives with scan errors are ten times more likely to fail than the group with no errors. This effect is also noticed when we further break down the groups by disk model. From Figure 8 we see a drastic and quick decrease in survival probability after the first scan error (left graph). A little over 70% of the drives survive the first 8 months after their first scan error. The dashed lines represent the 95% confidence interval. The middle plot in Figure 8 separates the population in four age groups (in months), and shows an effect that is not visible in the AFR plots. It appears that scan errors affect the survival probability of young drives more dramatically very soon after the first scan error occurs, but after the first month the curve flattens out. Older drives, however, continue to see a steady decline in survival probability throughout the 8-month period. This behavior could be another manifestation of infant mortality phenomenon. The right graph in figure 8 looks at the effect of multiple scan errors. While drives with one error are more likely to fail than those with none, drives with multiple errors fail even more quickly. Figure 6: AFR for scan errors. Figure 7: AFR for reallocation counts. Figure 8: Impact of scan errors on survival probability. Left figure shows aggregate survival probability for all drives after first scan error. Middle figure breaks down survival probability per drive ages in months. Right figure breaks down drives by their number of scan errors. The critical threshold analysis confirms what the charts visually imply: the critical threshold for scan errors is one. After the first scan error, drives are 39 times more likely to fail within 60 days than drives without scan errors. 3.5.2 Reallocation Counts When the drive’s logic believes that a sector is damaged (typically as a result of recurring soft errors or a hard error) it can remap the faulty sector number to a new physical sector drawn from a pool of spares. Reallocation counts reflect the number of times this has happened, and is seen as an indication of drive surface wear. About 9% of our population has reallocation counts greater than zero. Although some of our drive models show higher absolute values than others, the trends we observe are similar across all models. As with scan errors, the presence of reallocations seems to have a consistent impact on AFR for all age groups (Figure 7), even if slightly less pronounced. Drives with one or more reallocations do fail more often than those with none. The average impact on AFR appears to be between a factor of 3-6x. Figure 11 shows the survival probability after the first reallocation. We truncate the graph to 8.5 months, due to a drastic decrease in the confidence levels after that point. In general, the left graph shows, about 85% of the drives survive past 8 months after the first reallocation. The effect is more pronounced (middle graph) for drives in the age ranges [10,20) and [20, 60] months, while newer drives in the range [0,5) months suffer more than their next generation. This could again be due to infant mortality effects, although it appears to be less drastic in this case than for scan errors. After their first reallocation, drives are over 14 times more likely to fail within 60 days than drives without reallocation counts, making the critical threshold for this parameter also one. Figure 9: AFR for offline reallocation count. Figure 10: AFR for probational count. Figure 11: Impact of reallocation count values on survival probability. Left figure shows aggregate survival probability for all drives after first reallocation. Middle figure breaks down survival probability per drive ages in months. Right figure breaks down drives by their number of reallocations. 3.5.3 Offline Reallocations Offline reallocations are defined as a subset of the reallocation counts studied previously, in which only reallocated sectors found during background scrubbing are counted. In other words, it should exclude sectors that are reallocated as a result of errors found during actual I/O operations. Although this definition mostly holds, we see evidence that certain disk models do not implement this definition. For instance, some models show more offline reallocations than total reallocations. Since the impact of offline reallocations appears to be significant and not identical to that of total reallocations, we decided to present it separately (Figure 9). About 4% of our population shows non-zero values for offline reallocations, and they tend to be concentrated on a particular subset of drive models. Overall, the effects on survival probability of offline reallocation seem to be more drastic than those of total reallocations, as seen in Figure 12 (as before, some curves are clipped at 8 months because our data for those points were not within high confidence intervals). Drives in the older age groups appear to be more highly affected by it, although we are unable to attribute this effect to age given the different model mixes in the various age groups. After the first offline reallocation, drives have over 21 times higher chances of failure within 60 days than drives without offline reallocations; an effect that is again more drastic than total reallocations. Our data suggest that, although offline reallocations could be an important parameter affecting failures, it is particularly important to interpret trends in these values within specific models, since there is some evidence that different drive models may classify reallocations differ- ently. 3.5.4 Probational Counts Disk drives put suspect bad sectors “on probation” until they either fail permanently and are reallocated or continue to work without problems. Probational counts, Figure 12: Impact of offline reallocation on survival probability. Left figure shows aggregate survival probability for all drives after first offline reallocation. Middle figure breaks down survival probability per drive ages in months. Right figure breaks down drives by their number offline reallocation. Figure 13: Impact of probational count values on survival probability. Left figure shows aggregate survival probability for all drives after first probational count. Middle figure breaks down survival probability per drive ages in months. Right figure breaks down drives by their number of probational counts. therefore, can be seen as a softer error indication. It could provide earlier warning of possible problems but might also be a weaker signal, in that sectors on probation may indeed never be reallocated. About 2% of our drives had non-zero probational count values. We note that this number is lower than both online and offline reallocation counts, likely indicating that sectors may be removed from probation after further observation of their behavior. Once more, the distribution of drives with non-zero probational counts are somewhat skewed towards a subset of disk drive models. Figures 10 and 13 show that probational count trends are generally similar to those observed for offline reallocations, with age group being somewhat less pronounced. The critical threshold for probational counts is also one: after the first event, drives are 16 times more likely to fail within 60 days than drives with zero probational counts. 3.5.5 Miscellaneous Signals In addition to the SMART parameters described in the previous sections, which we have found to most closely impact failure rates, we have also studied several other parameters from the SMART set as well as other environmental factors. Here we briefly mention our relevant findings for some of those parameters. Seek Errors. Seek errors occur when a disk drive fails to properly track a sector and needs to wait for another revolution to read or write from or to a sector. Drives report it as a rate, and it is meant to be used in combination with model-specific thresholds. When examining our population, we find that seek errors are widespread within drives of one manufacturer only, while others are more conservative in showing this kind of errors. For this one manufacturer, the trend in seek errors is not clear, changing from one vintage to another. For other manufacturers, there is no correlation between failure rates and seek errors. CRC Errors. Cyclic redundancy check (CRC) errors are detected during data transmission between the physical media and the interface. Although we do observe some correlation between higher CRC counts and failures, those effects are somewhat less pronounced. CRC errors are less indicative of drive failures than that of cables and connectors. About 2% of our population had CRC errors. Power Cycles. The power cycles indicator counts the number of times a drive is powered up and down. In a server-class deployment, in which drives are powered continuously, we do not expect to reach high enough power cycle counts to see any effects on failure rates. Our results find that for drives aged up to two years, this is true, there is no significant correlation between failures and high power cycles count. But for drives 3 years and older, higher power cycle counts can increase the absolute failure rate by over 2%. We believe this is due more to our population mix than to aging effects. Moreover, this correlation could be the effect (not the cause) of troubled machines that require many repair iterations and thus many power cycles to be fixed. Calibration Retries. We were unable to reach a consistent and clear definition of this SMART parameter from public documents as well as consultations with some of the disk manufacturers. Nevertheless, our observations do not indicate that this is a particularly useful parameter for the goals of this study. Under 0.3% of our drives have calibration retries, and of that group only about 2% have failed, making this a very weak and imprecise signal when compared with other SMART parameters. Spin Retries. Counts the number of retries when the drive is attempting to spin up. We did not register a single count within our entire population. Power-on hours Although we do not dispute that power-on hours might have an effect on drive lifetime, it happens that in our deployment the age of the drive is an excellent approximation for that parameter, given that our drives remain powered on for most of their life time. Vibration This is not a parameter that is part of the SMART set, but it is one that is of general concern in designing drive enclosures as most manufacturers describe how vibration can affect both performance and reliability of disk drives. Unfortunately we do not have sensor information to measure this effect directly for drives in service. We attempted to indirectly infer vibration effects by considering the differences in failure rates between systems with a single drive and those with multiple drives, but those experiments were not controlled enough for other possible factors to allow us to reach any conclusions. 3.5.6 Predictive Power of SMART Parameters Given how strongly correlated some SMART parameters were found to be with higher failure rates, we were hopeful that accurate predictive failure models based on SMART signals could be created. Predictive models are very useful in that they can reduce service disruption due to failed components and allow for more efficient scheduled maintenance processes to replace the less efficient (and reactive) repairs procedures. In fact, one of the main motivations for SMART was to provide enough insight into disk drive behavior to enable such models to be built. After our initial attempts to derive such models yielded relatively unimpressive results, we turned to the question of what might be the upper bound of the accuracy of any model based solely on SMART parameters. Our results are surprising, if not somewhat disappointing. Out of all failed drives, over 56% of them have no count in any of the four strong SMART signals, namely scan errors, reallocation count, offline reallocation, and probational count. In other words, models based only on those signals can never predict more than half of the failed drives. Figure 14 shows that even when we add all remaining SMART parameters (except temperature) we still find that over 36% of all failed drives had zero counts on all variables. This population includes seek error rates, which we have observed to be widespread in our population (> 72% of our drives have it) which further reduces the sample size of drives without any errors. It is difficult to add temperature to this analysis since despite it being reported as part of SMART there are no crisp thresholds that directly indicate errors. However, if we arbitrarily assume that spending more than 50% of the observed time above 40C is an indication of possible problem, and add those drives to the set of predictable failures, we still are left with about 36% of all drives with no failure signals at all. Actual useful models, which need to have small false-positive rates are in fact likely to do much worse than these limits might sug- gest. We conclude that it is unlikely that SMART data alone can be effectively used to build models that predict failures of individual drives. SMART parameters still appear to be useful in reasoning about the aggregate reliability of large disk populations, which is still very important for logistics and supply-chain planning. It is possible, however, that models that use parameters beyond those provided by SMART could achieve significantly better accuracies. For example, performance anomalies and other application or operating system signals could be useful in conjunction with SMART data to create more powerful models. We plan to explore this possibility in our future work. Figure 14: Percentage of failed drives with SMART errors. 4 Related Work Previous studies in this area generally fall into two categories: vendor (disk drive or storage appliance) technical papers and user experience studies. Disk vendors studies provide valuable insight into the electromechanical characteristics of disks and both modelbased and experimental data that suggests how several environmental factors and usage activities can affect device lifetime. Yang and Sun [21] and Cole [4] describe the processes and experimental setup used by Quantum and Seagate to test new units and the models that attempt to make long-term reliability predictions based on accelerated life tests of small populations. Power-on-hours, duty cycle, temperature are identified as the key deployment parameters that impact failure rates, each of them having the potential to double failure rates when going from nominal to extreme values. For example, Cole presents thermal de-rating models showing that MTBF could degrade by as much as 50% when going from operating temperatures of 30C to 40C. Cole’s report also presents yearly failure rates from Seagate’s warranty database, indicating a linear decrease in annual failure rates from 1.2% in the first year to 0.39% in the third (and last year of record). In our study, we did not find much correlation between failure rate and either elevated temperature or utilization. It is the most surprising result of our study. Our annualized failure rates were generally higher than those reported by vendors, and more consistent with other user experience studies. Shah and Elerath have written several papers based on the behavior of disk drives inside Network Appliance storage products [6, 7, 19]. They use a reliability database that includes field failure statistics as well as support logs, and their position as an appliance vendor enables them more control and visibility into actual deployments 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 confidentiality concerns), their work is useful in enabling a qualitative understanding of factors what affect disk drive reliability. For example, they comment that end-user failure rates can be as much as ten times higher than what the drive manufacturer might expect [7]; they report in [6] a strong experimental correlation between number of heads and higher failure rates (an effect that is also predicted by the models in [4]); and they observe that different failure mechanisms are at play at different phases of a drive lifetime [19]. Generally, our findings are in line with these re- sults. User experience studies may lack the depth of insight into the device inner workings that is possible in manufacturer reports, but they are essential in understanding device behavior in real-world deployments. Unfortunately, there are very few such studies to date, probably due to the large number of devices needed to observe statistically significant results and the complex infrastructure required to track failures and their contributing factors. Talagala and Patterson [20] perform a detailed error analysis of 368 SCSI disk drives over an eighteen month period, reporting a failure rate of 1.9%. Results on a larger number of desktop-class ATA drives under deployment at the Internet Archive are presented by Schwarz et al [17]. They report on a 2% failure rate for a population of 2489 disks during 2005, while mentioning that replacement rates have been as high as 6% in the past. Gray and van Ingen [9] cite observed failure rates ranging from 3.3-6% in two large web properties with 22,400 and 15,805 disks respectively. A recent study by Schroeder and Gibson [16] helps shed light into the statistical properties of disk drive failures. The study uses failure data from several large scale deployments, including a large number of SATA drives. They report a significant overestimation of mean time to failure by manufacturers and a lack of infant mortality effects. None of these user studies have attempted to correlate failures with SMART parameters or other environmental factors. We are aware of two groups that have attempted to correlate SMART parameters with failure statistics. Hughes et al [11, 13, 14] and Hamerly and Elkan [10]. The largest populations studied by these groups was of 3744 and 1934 drives and they derive failure models that achieve predictive rates as high as 30%, at false positive rates of about 0.2% (that false-positive rate corresponded to a number of drives between 20-43% of the drives that actually failed in their studies). Hughes et al. 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. Our study takes a next step towards a better understanding of disk drive failure characteristics by essentially combining some of the best characteristics of studies from vendor database analysis, namely population size, with the kind of visibility into a real-world deployment that is only possible with end-user data. 5 Conclusions In this study we report on the failure characteristics of consumer-grade disk drives. To our knowledge, the study is unprecedented in that it uses a much larger population size than has been previously reported and presents a comprehensive analysis of the correlation between failures and several parameters that are believed to affect disk lifetime. Such analysis is made possible by a new highly parallel health data collection and analysis infrastructure, and by the sheer size of our computing deployment. One of our key findings has been the lack of a consistent pattern of higher failure rates for higher temperature drives or for those drives at higher utilization levels. Such correlations have been repeatedly highlighted by previous studies, but we are unable to confirm them by observing our population. Although our data do not allow us to conclude that there is no such correlation, it provides strong evidence to suggest that other effects may be more prominent in affecting disk drive reliability in the context of a professionally managed data center deployment. Our results confirm the findings of previous smaller population studies that suggest that some of the SMART parameters are well-correlated with higher failure probabilities. We find, for example, that after their first scan error, drives are 39 times more likely to fail within 60 days than drives with no such errors. First errors in reallocations, offline reallocations, and probational counts are also strongly correlated to higher failure probabilities. Despite those strong correlations, we find that failure prediction models based on SMART parameters alone are likely to be severely limited in their prediction accuracy, given that a large fraction of our failed drives have shown no SMART error signals whatsoever. This result suggests that SMART models are more useful in predicting trends for large aggregate populations than for individual components. It also suggests that powerful predictive models need to make use of signals beyond those provided by SMART. Acknowledgments We wish to acknowledge the contribution of numerous Google colleagues, particularly in the Platforms and Hardware Operations teams, who made this study possible, directly or indirectly; among them: Xiaobo Fan, Greg Slaughter, Don Yang, Jeremy Kubica, Jim Winget, Caio Villela, Justin Moore, Henry Green, Taliver Heath, and Walt Drummond. We are also thankful to our shepherd, Mary Baker for comments and guidance. A special thanks to Urs H¨olzle for his extensive feedback on our drafts. References [1] The r project for statistical computing. http://www.r-project.org. [2] Dave Anderson, Jim Dykes, and Erik Riedel. More than an interface - scsi vs. ata. In Proceedings of the 2nd USENIX Conference on File and Storage Technologies (FAST’03), pages 245 – 257, February 2003. [3] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E. Gruber. Bigtable: A distributed storage system for structured data. In Proceedings of the 7th USENIX Symposium on Operating Systems Design and Implementation (OSDI’06), November 2006. [4] Gerry Cole. Estimating drive reliability in desktop computers and consumer electronics systems. Seagate Technology Paper TP-338.1, November 2000. [5] Jeffrey Dean and Sanjay Ghemawat. Mapreduce: Simplified data processing on large clusters. In Proceedings of the 6th USENIX Symposium on Operating Systems Design and Implementation (OSDI’04), pages 137 – 150, December 2004. [6] Jon G. Elerath and Sandeep Shah. Disk drive reliability case study: Dependence upon fly-height and quantity of heads. In Proceedings of the Annual Symposium on Reliability and Maintainability, January 2003. [7] Jon G. Elerath and Sandeep Shah. Server class disk drives: How reliable are they? In Proceedings of the Annual Symposium on Reliability and Maintainability, pages 151 – 156, January 2004. [8] Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung. The google file system. In Proceedings of the 19th ACM Symposium on Operating Systems Principles, pages 29 – 43, December 2003. [9] Jim Gray and Catherine van Ingen. Empirical measurements of disk failure rates and error rates. Technical Report MSR-TR-2005-166, December 2005. [10] Greg Hamerly and Charles Elkan. Bayesian approaches to failure prediction for disk drives. In Proceedings of the Eighteenth International Conference on Machine Learning (ICML’01), June 2001. [11] Gordon F. Hughes, Joseph F. Murray, Kenneth Kreutz-Delgado, and Charles Elkan. Improved disk-drive failure warnings. IEEE Transactions on Reliability, 51(3):350 – 357, September 2002. [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 using non-parametric statistical methods. Proceedings of ICANN/ICONIP, June 2003. [14] Joseph F. Murray, Gordon F. Hughes, and Kenneth Kreutz-Delgado. Machine learning methods for predicting failures in hard drives: A multipleinstance application. J. Mach. Learn. Res., 6:783– 816, 2005. [15] Rob Pike, Sean Dorward, Robert Griesemer, and Sean Quinlan. Interpreting the data: Parallel analysis with sawzall. Scientific Programming Journal, Special Issue on Grids and Worldwide Computing Programming Models and Infrastructure, 13(4):227 – 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 Conference on Mass Storage Systems and Technologies, May 2006. [18] Sandeep Shah and Jon G. Elerath. Disk drive vintage and its effect on reliability. In Proceedings of the Annual Symposium on Reliability and Maintainability, pages 163 – 167, January 2004. [19] Sandeep Shah and Jon G. Elerath. Reliability analysis of disk drive failure mechanisms. In Proceedings of the Annual Symposium on Reliability and Maintainability, pages 226 – 231, January 2005. [20] Nisha Talagala and David Patterson. An analysis of error behavior in a large storage system. Technical Report CSD-99-1042, University of California, Berkeley, February 1999. [21] Jimmy Yang and Feng-Bin Sun. A comprehensive review of hard-disk drive reliability. In Proceedings of the Annual Symposium on Reliability and Maintainability, pages 403 – 409, January 1999.