Hard Disk Internal Performance Factors
I refer to performance considerations that relate only or primarily to the capabilities of the hard disk drive itself as internal performance factors. In theory, these are not subject to any restrictions imposed by the interface or other parts of the system, which means they should be reasonably consistent and even "portable" from one system to another. These are really the basis of hard disk performance, since they dictate the theoretical maximums. Note that many of these performance factors are inter-related and some overlap or are based on others listed here.
Positioning and Transfer Performance Factors
The process of reading or writing to the hard disk really comprises two large steps, the work done by the drive itself (measured by internal performance factors) and the interface (external factors). The internal work of the drive can itself be thought of as two functions: finding the correct location on the drive, and then reading or writing the data. These are very different jobs, and two drives can be very similar in one regard but very different in another. They also depend on different design characteristics. I call these tasks positioning and transfer.
Both of these factors are important to overall performance, although if you read the literature and the numbers that people talk about, positioning metrics are probably more commonly discussed than transfer measurements. You might be fooled by this into thinking they are more important, but often they are not--they are just simpler to explain in many cases, or people are used to using them to compare drives.
Which influences on performance are most important also depends on how you are using the device. If you are running a file server, the hard disk will be doing a lot of random accesses to files all over the disk, and positioning performance will be very important. If you are a single user doing multimedia editing where you need to read multi-megabyte consecutive files as fast as possible, data transfer is far more important than positioning speed.
Positioning performance factors (or characteristics with a direct and significant influence on positioning performance) discussed in this chapter include:
Transfer performance factors (or characteristics with a direct and significant influence on transfer performance) discussed in this chapter include:
Every performance factor is also directly influenced by a myriad of other design decisions that go into the creation of the drive.
Seek Time
The seek time of a hard disk measures the amount of time required for the read/write heads to move between tracks. This is one of the most commonly stated metrics for hard disks, and it is one of the most important positioning performance factors. However, using this metric to compare drives can be somewhat dangerous. To use it properly, we must figure out exactly what seek time means.
Switching between tracks requires the head actuator to move the head arms physically, which being a mechanical process, takes a specific amount of time. The amount of time to switch between two tracks depends on the distance between the tracks. However, there is a certain amount of overhead involved in track switching, so the relationship is not linear. It does not take double the time to switch from track 1 to track 3 that it does to switch from track 1 to track 2, much as a trip to the drug store 2 miles away does not take double the time of a trip to the grocery store 1 mile away, when you include the overhead of getting into the car, starting it, etc.
Seek time is normally expressed in milliseconds, with average seek times for most modern drives in the 8 to 12 ms range. In the modern PC, a millisecond is an enormous amount of time: your system memory has speed measured in nanoseconds, for example (one million times smaller). A 200 MHz processor can (theoretically) execute 200,000 instructions in a millisecond. Cutting the seek time of an average read instruction down from 12 ms to 8 ms then, can result in great system performance improvement, because the rest of the system is often sitting and waiting for the hard disk during this time.
There are problems, however, with using seek time to compare hard disks. The first one is that there is no standardized way of reporting them. Most companies only specify a single seek time for their drive. This is obviously problematic, because there is no single number that expresses the seek time for an entire drive. The time to seek from one track to another one depends on the distance between them. Many manufacturers seek to avoid this problem by defining and quoting an "average" seek time. In fairness, this is a good way of representing the seek time for overall use, but the definition of what seek patterns represent "average" is left basically up to each manufacturer to decide. In many ways a truly random access pattern is not realistic, since most users don't do truly random access to their disks.
Other manufacturers provide more information about the seek time of their drives than just the average. Here are some more detailed seek time specifications that you can use to do a much more complete comparison of two drives:
A second problem is that so many drives have near-identical seek times. At any given time, the leading-edge drives are usually within 1 ms or less of each other in terms of average seek time, which doesn't leave much to distinguish them. To make matters worse, some manufacturers don't even quote a true average seek time, saying just "less than 11 ms" for example. This doesn't help you terribly, because in some ways this really means "we're not entirely sure". (At least it tells you what they think of seek time as a metric.)
Finally, while seek time is an important component of overall hard disk performance, it is only one component. There are situations in which it is very unimportant as an overall factor. If you are doing work that involves reading large blocks of contiguous data from the disk, average seek time is much less important to you than the drive's transfer rate and cylinder switch time, for example. Don't be fooled by drive manufacturers and others that place so much importance on seek time that they even refer to their drives as a "10 ms hard disk" when their average seek time is this value.
Latency
The hard disk platters are spinning around at high speed, and the spin speed is not synchronized to the process that moves the read/write heads to the correct cylinder on a random access of any sector of the hard disk. Therefore, at the time that the heads arrive at the correct cylinder, the actual sector that is needed may be anywhere. Latency is the time that the drive must wait for the correct sector to come around to where the read/write heads are waiting for it.
Conceptually, latency is rather simple to understand; it is also easy to calculate. The faster the disk is spinning, the quicker the correct sector will rotate under the heads, and the lower latency will be. On average, latency will be half the time it takes for a full rotation of the disk. This table shows the latency for a range of typical hard disk spindle speeds:
Spindle Speed (RPM) | Average Latency (Half Rotation) (ms) | Worst-Case Latency (Full Rotation) (ms) |
3,600 | 8.3 | 16.7 |
4,500 | 6.7 | 13.3 |
5,200 | 5.8 | 11.5 |
5,400 | 5.6 | 11.1 |
6,300 | 4.8 | 9.5 |
7,200 | 4.2 | 8.3 |
10,000 | 3.0 | 6.0 |
As with seek time, latency figures are in milliseconds, which are big numbers when dealing with computer system performance. Reducing latency figures greatly improves the performance of the system, particularly when dealing with access to random files on the disk. This is why newer and more expensive hard disks continue to push spindle speeds higher and higher. Even most mainstream disks now rotate at 5,200 or 5,400 RPM. You can see that a 5,400 RPM drive gets two-thirds of the latency improvement of a 7,200 RPM drive, compared to the old standard of 3,600 (which all PC hard disk drives used for many years.)
Again, as with seek times, latency is most relevant only to certain types of accesses. For multiple, frequent reads of random sectors on the disk, it is an important performance-limiting factor. For reading large continuous blocks of data, latency is a relatively minor factor because it will only happen while waiting to read the first sector of a file. The use of cylinder and head skewing on modern drives is intentionally designed to reduce latency considerations when switching between consecutive heads or cylinders.
Access Time
Positioning on a random read or write of the disk requires moving the heads to the correct cylinder and then waiting for the correct sector to rotate under the heads. These are measured by seek time and average latency respectively. It is also possible to combine these two figures into a composite number; this is called access time and is measured in milliseconds. This metric is not used nearly as frequently as seek time for hard disks. It is much more commonly used in assessing CD-ROM drives
Head Switch Time
Each cylinder contains a number of tracks, each accessible by one of the heads on the drive. To improve efficiency, the drive will normally use all of the tracks in a cylinder when doing a sequential read or write, because this saves the time required to physically move the heads to a new cylinder. Switching between heads is a purely electronic process instead of a mechanical one. However, switching between heads within a cylinder still requires a certain amount of time, called the head switch time. This is usually less than the track switch time.
The head switch time of the drive is often not specified. It is, compared to the other performance metrics, of somewhat less significance, but it should be realized that a drive with multiple platters doing a large block read or write will switch between heads much more often than it will switch between tracks.
Track Switch Time
The track switch time, also called cylinder switch time, measures in milliseconds the amount of time required to move the read/write heads from one cylinder to an adjacent one. This is a mechanical process that involves using the actuator to physically move the read/write heads. Track switch time is really a special case of seek time, where the seek is being done to an adjacent track. See in that description for some more information, some of which is relevant to track switch time as well.
Track switch time is reasonably important because switches to adjacent tracks occur much more frequently than random seeks, when processing larger files. Even during a continuous transfer of large size, the heads must be moved from track to track. An average cylinder on even a modern high-density disk contains less than 1 megabyte of data, which means that a multi-megabyte read or write will involve many cylinder switches.
Internal Data Transfer Rate
Since the obvious objective in using a hard disk is to transfer data to the hard drive and onto the disks, or off the disks and out of the drive, the rate of data transfer is of paramount importance. It is my personal opinion that this is an underrated performance metric, especially compared to the much more commonly stated metrics of seek time and interface transfer rate.
First, a word about terminology. Transfer rates are confusing in part because of the phrase "transfer rate" can mean so many different things. What we are talking about here is the rate at which the hard disk can physically read data from the surface of the platter and transfer it to the internal drive cache or read buffer, ready for sending over the interface to the system. This is the drive's internal data transfer rate. This is opposed to the speed that the data can then be sent from the buffer over the interface to the system, which is the external data transfer rate.
The internal data transfer rate, which is the real rate that data can be read from the disk, is often called the sustained transfer rate, while the external rate is called the peak or burst transfer rate. The reason for these terms is that the external rate is usually much higher than the internal rate. So the drive can burst data over the interface at the higher rate when it finds the data requested already in the buffer. But the buffer is quite small compared to the size of the disk (less than 1 MB in most cases), so for a sustained read of any reasonable size, the platters themselves must be accessed, and the overall data transfer rate will drop down to only whatever the drive can handle internally.
To make matters even more confusing (sorry!) even the internal data transfer rate has more than one value. No disk can maintain even its sustained transfer rate over a prolonged period of time, because this rate is only produced under ideal conditions: reading a small number of consecutive sectors over the fastest part of the disk. In the "real world", reading data involves seeking to different parts of the disk, using different heads on different platters, etc., so that a read of a 1 MB file from your disk will never proceed at the full stated maximum sustained transfer rate (although it will be much closer than to the burst transfer rate).
Calculating the data transfer rate is reasonably simple, provided you know the true specifications of the drive; figuring the transfer rate will show you what design parameters have an impact on this performance measure. The transfer rate is a measure of the amount of data that can be accessed over a period of time. So we need to know how much data is able to pass under the read/write heads in one second. This is dependent on the density of the data (how tightly packed the data is into each linear inch of disk track), and also how fast the disk is spinning. The density of the data can be calculated easily if we know how many sectors are on the track, since we know how many bytes there are in a sector. The speed of the disk is calculated in RPM, so we divide it by 60 to get revolutions per second. This gives us a calculation of the data transfer rate in megabits per second as follows (to get the result in megabytes per second, simply divide by 8):
Note: You need the true physical geometry here; the logical BIOS setup
parameters will give incorrect results. If the geometry you are using says the disk has 63
sectors per track and 16 heads, chances are very high that you are looking at the logical
BIOS geometry.
Recall that modern disk drives use zoned bit recording, which means that the inner tracks have fewer sectors per track than the outer ones. Since the data transfer rate is directly proportional to the number of sectors per track, this means the data transfer rate for the outside tracks of a disk can be as much as double the data transfer rate for the inside tracks! Let's take as an example the Quantum Fireball TM drive whose zones and quoted transfer rates we saw in the section on ZBR. The outermost zone uses 232 sectors per track. The drive is spinning at 5400 RPM, so this yields a total of 85.5 Mbits/s or 10.7 MB/s. This is equal to Quantum's stated maximum data transfer rate for the drive.
You may notice that in the table in the discussion of ZBR, there is a stated data transfer rate for the outermost zone of 92.9 Mbits/s, which is substantially higher than the 85.5 number we calculated. The reason is that we only looked at the "real" data that was being read from the drive, 512 bytes per sector. Each sector on this drive actually holds 540 bytes, because 28 bytes are used for ECC. If we replace the 512 in the formula with 540, we get 90.2 Mbits/s. The remaining difference is probably from other overhead associated with the control structures of the drive.
When looking at quoted or measured data transfer rates, it's important to take several things into account. There are various assumptions and tricks that various manufacturers use in different ways to help paint their products in the best possible light. There are also problems with how some benchmarks that claim to measure transfer rates really do it:
Spindle Speed
The drive's spindle speed is not a performance measure per se, but it is a characteristic of the drive that is very important in determining how well the drive will perform. In fact, it is one of the better "single numbers" in terms of its reliability in being an indicator of overall hard disk performance: in virtually every case a hard disk that runs at 7,200 RPM is going to offer better performance than one running at 5,400 RPM. In this regard, it is in my opinion far more useful than knowing the rated seek time of the drive, for example. It is also a readily available statistic and easy to determine for any drive.
The reason why spindle speed is so important is that it directly impacts both positioning time and data transfer. The impact on positioning time is through reduction inlatency; the impact on data transfer rate is direct because the heads can only read from the disk at the rate they spin past them. This means that faster spindle speed drives will have improved performance regardless of whether they are used for many small, random accesses, or for streaming large contiguous blocks from the disk.
Areal Density
The areal density of the drive, or more correctly of its platters, measures how many bits can be packed into each square inch of surface area. Areal density is discussed in detail here. Like spindle speed, areal density is an important characteristic that determines overall performance. It is, however, somewhat more difficult to determine. Most manufacturers don't talk a great deal about areal density, and usually only mention it (if at all) in their detailed specifications for a drive.
Areal density is important because, much like spindle speed, it impacts on both positioning time and data transfer. Areal density consists of two components, and each of these impacts on one of these areas:
Note that some drives improve areal density by increasing only one of these factors, while others increase them both. Determining actual areal density figures for drives can be difficult, but you can use the readily available statistics of various drives to do rough sketch comparisons quite easily. One useful measurement that I have employed to compare the areal density of drives is capacity per surface. It is calculated as follows:
Capacity per surface = Total formatted capacity / Number of read-write heads
The number of read-write heads is usually equal to double the number of platters, but not always, since some drives use only a single head on one of the platters in the drive. Note that you need to know the real physical number of heads for the drive, not the logical geometry (which will usually state 16 heads regardless of the real number).
It is easy to compare the areal density of drives using this technique, and this tells you a lot about the expected performance of the drives. Let's take as an example a series of Western Digital Caviar hard drives, which have a convenient model numbering that tells us their approximate capacity and number of platters instantly (although not the number of read/write heads, which is double the number of platters with two exceptions):
Caviar Model Number | Approximate Capacity (Decimal MB) | Read/Write Heads | Approximate Capacity per Surface (MB) |
AC2850 | 850 | 4 | 212 |
AC31000 | 1,000 | 6 | 167 |
AC21000 | 1,000 | 4 | 250 |
AC21200 | 1,200 | 4 | 300 |
AC31600 | 1,600 | 6 | 267 |
AC21600 | 1,600 | 4 | 400 |
AC32100 | 2,100 | 5 | 420 |
AC22100 | 2,100 | 4 | 525 |
AC32500 | 2,500 | 5 | 500 |
AC33100 | 3,100 | 6 | 517 |
AC34000 | 4,000 | 6 | 667 |
As you can see, the density per surface increases with the newer, larger capacity drives. But you can also see that some drives have the same capacity, but much higher density. This is because they are newer generation drives, and the technology has advanced to allow them to fit the same capacity in a smaller number of surfaces. An AC21000 is going to provide faster data transfer than an AC31000 for this reason.
There is one other factor here as well. Generally speaking--not always, but usually--drives with higher density are newer, and therefore they generally have design improvements in other areas: newer controller electronics, etc., so they are even more likely to be better than the older drive overall (assuming all else is equal of course).
Hard Disk External Performance Factors
Performance factors that are solely a function of the capabilities of the hard disk drive itself are the internal performance factors. However, there are also external performance factors to be considered, which relate to how the hard disk relates to other components in the system. The most important factor here is the speed of the interface between the hard drive and the rest of the PC, which can be a bottleneck to peak performance.
This section discusses external factors that affect the performance of a hard disk in the "real world" of a system. Note that many of these factors are so external that they don't relate directly to the drive itself, but rather the rest of your system. This means that these factors can be different for the same drive in a different system.
Interface Speed and the External Data Transfer Rate
The objective of using a hard disk is to transfer data to and from the disks. This involves two steps: the internal part is the process of actually reading or writing the disk platters, and the external part is moving the data from the inside of the drive out to the system, or vice-versa. The internal transfer rate of the drive is discussed in this section, while the external data transfer rate is discussed here.
Transfer rates are confusing in part because of the many different types that are often discussed. The external transfer rate is also sometimes called the interface transfer rate or host transfer rate, because it is the speed of transfer over the interface between the hard disk and the rest of the PC. The external and internal transfer rates are often called the burst and sustained transfer rates, respectively. The external rate is called burst because it can be much higher than the sustained rate, for short transfers. The sustained rate is so named because when you are dealing with a large transfer, the limiting factor is how fast the disk's internal mechanisms can operate.
The external transfer rate is the speed at which data can be exchanged between the system memory and the internal buffer or cache built into the drive. This is usually faster than the internal rate because it is a purely electronic operation, which is much faster than the mechanical operations involved in accessing the physical disk platters themselves. This is in fact a major reason why modern disks have an internal buffer.
The external transfer rate is dictated primarily by the type of interface used, and themode that the interface operates in. Support for a given mode has two requirements: the drive itself must support it, and the system (usually meaning the system BIOS and chipset) must support it as well. Only one or the other does absolutely no good. Support for the higher transfer modes also usually requires that the interface be over a high-speed system bus such as VLB or PCI, which most today are. See also the section on system issues.
The two most popular hard disk interfaces used today are SCSI and IDE/ATA (and enhancements of each). IDE uses two types of transfer modes: PIO and DMA. I'm not going to get into all the details on interfaces and transfer modes here, because I have another large section devoted entirely to them. I did want to compare the theoretical maximum transfer rates of the different interface options, so I've "mashed" them together in the table below to show you a comparative summary:
Interface | Mode | Theoretical Transfer Rate (MB/s) |
IDE/ATA | Single Word DMA 0 | 2.1 |
IDE/ATA | PIO 0 | 3.3 |
IDE/ATA | Single Word DMA 1 | 4.2 |
IDE/ATA | Multiword DMA 0 | 4.2 |
Standard SCSI | -- | 5.0 |
IDE/ATA | PIO 1 | 5.2 |
IDE/ATA | PIO 2 | 8.3 |
IDE/ATA | Single Word DMA 2 | 8.3 |
Wide SCSI | -- | 10.0 |
Fast SCSI | -- | 10.0 |
EIDE/ATA-2 | PIO 3 | 11.1 |
EIDE/ATA-2 | Multiword DMA 1 | 13.3 |
EIDE/ATA-2 | PIO 4 | 16.6 |
EIDE/ATA-2 | Multiword DMA 2 | 16.6 |
Fast Wide SCSI | -- | 20.0 |
Ultra SCSI | -- | 20.0 |
Ultra ATA | Multiword DMA 3 (DMA-33) | 33.3 |
Ultra Wide SCSI | -- | 40.0 |
There are a boatload of caveats that go with these numbers, and you really need to understand the interfaces themselves to know what they are. Obviously these are theoretical maximums that don't take into account command processing overhead etc. Also see the comparison of SCSI and IDE for more issues of relevance to interface performance.
For high performance, it is important that the external (burst) transfer rate be higher than the internal (sustained) transfer rate of the drive. If it isn't, then the drive is not being used to its maximum potential and the interface should really be upgraded. The most popular hard disk interfaces used in modern PCs today are EIDE or ATA-2 drives running with PIO mode 4 or multiword DMA mode 2, meaning a theoretical maximum of 16.6 MB/s. This exceeds the internal transfer rate of any hard disk on the market today.
Once you get above the internal transfer rate, the extra "room" is only of value if whatever you are trying to read from the disk is already in the buffer. If the hard disk itself can only do sustained reads of 7 MB/s, then over time, the extra 9.6 MB/s of the 16.6 MB/s interface speed does not do a lot for you. However, having that extra slack is of use in making sure peak transfers have the capacity they need. Many hard disk manufacturers try to trick their buyers by putting the 16.6 MB/s speed of the interface in big letters in their ads, but as you know now, this is not the most important number. The newest drives support Ultra ATA and its 33.3 Mb/s interface speed, which makes the confusion even worse because virtually no systems are going to exploit even half of that capacity right now.
There is also another consideration: the number of devices that are sharing the interface. This is particularly a concern with SCSI, which allows for many devices on a bus (IDE/ATA and enhancements allow just two per channel). If you are using four hard disks on a SCSI bus in a server that is handling many simultaneous requests, and each drive has an internal transfer rate of 8 MB/s, that 40 MB/s for Ultra Wide SCSI will probably, at many points in time, be in full use.
System Factors
As I probably reiterate in too many places on this site, the components in a PC system are interrelated, and affect each others performance in many ways. This makes it difficult to measure the performance of any component in isolation. Some tests are better able than others to isolate the component being tested, but it also depends on the component. Hard disks are virtually impossible to completely isolate from the rest of the system, because every access of the hard disk involves a transfer through the main processing subsystems.
The following are the main parts of the system directly influence the performance of the hard disk subsystem:
The bottom line of all this is that comparing drives can only be fairly done using a comparative benchmark. In my opinion, you cannot get a system-independent single score that gives a fair assessment of a hard drive's performance.
Command Overhead and Multiple Device Considerations
When looking at interface specifications or raw transfer rates, don't forget that each transfer from the hard disk requires that a set of commands be sent to the hard disk to select what will be sent and to control the transfer. These take a certain amount of time, and this command overhead can and will reduce performance. In comparing the SCSI and IDE/ATA interfaces, command overhead is an important consideration. SCSI is a much more intelligent and capable bus, but it is also more complex, which means more work must be done to set up a transfer. This means that SCSI can be slower than IDE/ATA in a single-user, single-tasking environment, even though it can be much faster and more capable in a machine that is supporting multiple users or multiple devices on the same bus. SCSI shines when you need to use multiple devices on a single bus, where IDE/ATA starts to become cumbersome.
Disk Caching
The process of caching describes the use of buffers to separate operations that differ significantly in speed, so the fast one is not held up by the slower one. In a system there are many levels of caching that are used to allow different-speed components to run unimpeded; in the disk subsystem there are generally two levels. The disk drive's logic board contains an integral cache. This cache is used to separate the internal mechanical read/write operations from transfers over the bus, and to hold recently accessed data. A larger cache will result in improved performance by cutting down on the required number of physical seeks and transfers on the platters themselves. Smarter caching algorithms can have the same effect.
In addition to this hardware caching, most operating systems use software disk caching. Since the system memory is many orders of magnitude faster than the hard disk, a small area of system memory (usually one to a few megabytes) is set aside to buffer requests to the hard disk. When the disk is read, the data are stored in this cache in case they are needed again in the near future (which they often are). If they are needed again, they can be supplied from the cache memory instead of requiring another read of the hard disk. Again here, increasing the size of the cache improves performance--to a point. If you increase it too much, you will run out of usable memory for programs and data, and the system will be forced to rely on much slower virtual memory. In this case your use of memory as virtual disk is causing the system to also need to use your disk as virtual memory, defeating your original intent!
Some SCSI host adapters add an additional level of cache on the controller itself (sometimes called caching controllers). This cache can be several megabytes in size and logically sits between any system disk cache and any cache on the hard disk itself. Again, this improves performance by reducing the number of accesses required to the disk. In this case when the system tries to read data from the hard disk, the controller will intercept the request and if it is in the cache, satisfy it from there instead of going to the hard disk. This both improves system speed greatly and also cuts down on traffic on the SCSI bus.
File System Performance Factors
There are several factors related to how the disk is logically structured and the file system set up and maintained, that can have a tangible effect on performance. These are basically independent of the hard disk and will have similar impacts on any hard disk. See the section on the file system for much more information on these issues:
Redundant Arrays of Inexpensive Disk (RAID)
Many higher-end systems, especially servers, now employ a new technology called RAID. This system allows for great improvements in both reliability and performance. The idea is to store data on multiple disk drives running in parallel. The primary motivation in most cases is reliability, and for this reason RAID is discussed in detail here. From a performance standpoint, most RAID levels improve performance by allowing multiple accesses to happen simultaneously, and also by using algorithms that reduce seek time and latency by taking advantage of having multiple drives at their disposal. The exact performance impact depends on the level of RAID used; some improve read performance at the expense of write performance, for example.
Mean Time Between Failures (MTBF)
One of the most commonly used statistics related to drive reliability is mean time between failures or MTBF. This value, usually measured in hours, is meant to represent the average amount of time that will pass between random failures on a drive of this type. It is usually in the range of 200,000 to 500,000 hours for modern drives.
This number is very often misinterpreted and misused. Usually, it goes like this: "Gee, a year contains 8,766 hours. That means my 500,000 MTBF drive should last 57 years." Then, amusingly, one of two things happens: either the person actually thinks the drive will last 57 years, or the opposite: they realize this is crazy and so they write off the entire MTBF figure as "obvious exaggeration from the manufacturer and therefore useless". The real answer of course is neither.
It is obviously impossible for any individual hard disk to be tested to anywhere near the amount of time required to provide a MTBF factor near 500,000. MTBF is meant to be an estimated average, based on testing done on many hard disks over a smaller period of time. It should obviously be recognized that these are averages and that they are computed by the manufacturer, but they do have value when used properly.
The MTBF figure is intended to be used in conjunction with the useful service life of the drive, the typical amount of time before the drive enters the period where failures due to component wearout increase. If the MTBF is 500,000 hours and the service life is five years, this means that the drive is supposed to last for five years, and that of a large group of drives operating within this timeframe, on average they will accumulate 500,000 of total run time between failures. Or, you can think of it this way: if you used one of these drives and replaced it every five years with another identical one, in theory it should last 57 years before failing, on average (though I somehow doubt we'll be using 1 to 10 GB spinning-platter hard disk drives in the year 2050. )
Overall, MTBF is a "reasonably interesting" reliability statistic--not something totally useless, but definitely something to be taken with a grain of salt. I personally view the drive's warranty and stated service life to be more indicative of what the manufacturer really thinks of the drive. I personally would rather buy a hard disk with a stated service life of five years and a warranty of three years, than one with a service life of three years and warranty of two years, even if the former has an MTBF of 300,000 hours and the latter one of 500,000 hours.
Self-Monitoring Analysis and Reporting Technology (SMART)
In an effort to help users avoid data loss, some drive manufacturers are now incorporating logic into their drives that acts as an "early warning system" for pending drive problems. This system is called Self-Monitoring Analysis and Reporting Technology or SMART. The hard disk's integrated controller monitors various aspects of its own performance and makes available status information to any software that wants to probe the drive and look at it. One popular program that will monitor this status is Symantec's Norton Utilities suite.
SMART works by monitoring certain performance characteristics inside the disk, and looking for trends that indicate a gradual worsening in reliability that could indicate an imminent failure. The actual way that SMART is implemented is usually a function of what the manufacturer wants to do, and the specifics of the drive itself. One example of what SMART could do is that it could monitor the average number of ECC error corrections it must perform per sector read from the disk. Based on statistical knowledge and examining trends over time, it could be possible to predict the formation of bad sectors or total drive failure.
The value of SMART is still questionable. The problem with it is that it really cannot detect many types of drive failures. Also, consider that a drive showing up with a SMART code indicating pending failure during the warranty period of the drive, will mean an expensive replacement for the manufacturer. For this reason, I think it likely that they will make the SMART routines pretty conservative, to make sure that hard disks that don't need replacing aren't. This will limit the usefulness of the technique compared to what it would be if they were more aggressive in detecting these failures.
