Once you install it in … [Read More When it comes to browsing the internet and downloading files, security is the absolute priority. But once you're safe, there's also some other things … [Read More The structure of older hard drives i. A cylinder is formed while all drive heads are in the same position on the disk. The tracks, stacked on top of each other form a cylinder. This scheme is slowly being eliminated with modern hard drives. All new disks use a translation factor to make their actual hardware layout appear continuous, as this is the way that operating systems from Windows 95 onward like to work.
To the operating system of a computer, tracks are logical rather than physical in structure, and are established when the disk is low-level formatted. Tracks are numbered, starting at 0 the outermost edge of the disk , and going up to the highest numbered track, typically , close to the center. Similarly, there are 1, cylinders numbered from 0 to on a hard disk. The stack of platters rotate at a constant speed. The drive head, while positioned close to the center of the disk reads from a surface that is passing by more slowly than the surface at the outer edges of the disk.
To compensate for this physical difference, tracks near the outside of the disk are less-densely populated with data than the tracks near the center of the disk. The result of the different data density is that the same amount of data can be read over the same period of time, from any drive head position. The disk space is filled with data according to a standard plan.
One side of one platter contains space reserved for hardware track-positioning information and is not available to the operating system. Fragmentation is a problem when the file system must search several different locations to find all the pieces of the file you want to read. The search causes a delay before the file is retrieved. A larger cluster size reduces the potential for fragmentation, but increases the likelihood that clusters will have unused space.
Using clusters larger than one sector reduces fragmentation, and reduces the amount of disk space needed to store the information about the used and unused areas on the disk. The stack of platters rotate at a constant speed.
The drive head, while positioned close to the center of the disk reads from a surface that is passing by more slowly than the surface at the outer edges of the disk. To compensate for this physical difference, tracks near the outside of the disk are less-densely populated with data than the tracks near the center of the disk.
The result of the different data density is that the same amount of data can be read over the same period of time, from any drive head position. The disk space is filled with data according to a standard plan. One side of one platter contains space reserved for hardware track-positioning information and is not available to the operating system. Thus, a disk assembly containing two platters has three sides available for data.
Track-positioning data is written to the disk during assembly at the factory. The system disk controller reads this data to place the drive heads in the correct sector position. Data Recovery. Data Utility. The graph below plots the number of reads per second IOPS , which is the reciprocal of the access time when there are no parallel accesses no queueing.
IOPS increases when a smaller fraction of the disk is used. Short stroking a hard drive gives modest performance improvements by sacrificing a large amount of storage space. The plot shows two scenarios: One where the usable region is at the beginning of the disk the usual case when short stroking because the beginning of the disk is faster , and the other where the usable region is at the end of the disk the slowest region.
Because seek time is highly non-linear with distance short seeks still have a high cost , the improvement in random IOPS is small unless only a very tiny fraction of the disk is used.
Now that flash-based solid-state drives are common, short stroking a hard drive for performance makes little sense. As of July , a low-end SSD is about 3 to 4 times more expensive per capacity than hard drives, but offers times higher random IOPS and double the sequential throughput. Start and end are sector numbers. This measures the average access time for doing iterations random accesses within the region between start inclusive and end exclusive.
The algorithm reads random sectors. It measures access time, not seek time. Seek profile plotted on a polar axis. A seek profile is equivalent to finding the boundary between colour regions in the acess time plot. The seek time is the time it takes for the disk head to move to the destination track, but not including the rotational latency to wait for the target sector to rotate under the disk head.
I included head settling time in the seek time because settling is really just the final low-speed fine-tuning portion of the seek, and there is no way to measure settling time separately from the seek. Seek time provides information about the radial location of a given track on the disk.
Seeks of a longer physical distance usually take longer than seeks of a shorter physical distance, so seek time can be used to determine which tracks are physically closer to a reference location. I usually use sector 0 on the outer edge as the reference sector. While it is straightforward to measure access time, seek time requires finding the sector on the target track that has the lowest access time with near-zero rotational latency.
Nisha et al. My algorithm improves on speed by using a logarithmic search for the sector with the minimum access time starting with big steps that decrease exponentially the runtime is logarithmic in the number of sectors per track rather than linear.
This becomes particularly important when disks have many sectors per track. I tested disks with up to sectors per track, while Nisha et al.
The seek time profile is usually plotted with seek time on the vertical axis and seek distance number of tracks on the horizontal axis. The diagram to the right shows an example from a Toshiba P 3 TB drive. The seek time increases roughly linearly with distance for long seeks when the seek time is dominated by the head moving at a high constant speed and is non-linear for short seeks dominated by acceleration, deceleration, and settling time.
Another way to think of the seek time curve is that it is the boundary that separates colour regions in the access time plot. To demonstrate this, the two polar plots to the right plots the seek profile and access time data points for the same drive on the same polar axis for comparison. The seek profile is usually more useful when plotted in cartesian coordinates. Western Digital S The seek profile is unusual, indicating a strange mapping between track numbers and physical location.
This mapping seems quite irregular. Samsung F3. Even when completely zoomed out, four recording surfaces are visible. The two plots on the right are examples of slightly more complicated seek time profiles. These patterns indicate that tracks are not strictly placed from outside diameter to inside diameter, but that there are some later logical tracks that are physically closer to the outside lower seek time than some earlier tracks higher seek time.
These patterns give information on the track layout, which we will look at in the track layout section. The access time of a sector depends on its angular position. The algorithm to find seek time searches nearby earlier sectors to find a local minimum sector with near-zero rotational latency. The seek time is the access time for the sector on the target track that has the lowest access time near-zero rotational latency. This algorithm starts by measuring the access time including the rotational latency to an arbitrary target sector.
It then measures the access time for an earlier sector, chosen to be likely on the same track but with less rotational latency. This process is repeated until the access time suddenly increases, which occurs when the target sector arrives too soon under the head and we need to wait one more full revolution.
The step size is then halved and the search repeated until the sector with minimum access time is found. This implementation tries to mitigate this by attempting to search two adjacent tracks before reporting the seek time. Like the access time measurement, the seek time measurement sends read requests serially and thus includes the OS and disk controller overhead. I have not attempted to use this.
Example: The Cheetah 15K. Samsung F3: Full-stroke seek increases from Automatic Acoustic Management AAM is a method to allow the user to reduce audible noise from the hard drive in exchange for lower performance. Since most of the noise comes from moving the disk head during seeks, AAM essentially slows down seeks increasing seek time to reduce noise. Using the ability to measure seek time, we can see how AAM quiet mode affects seeks. The diagrams show a seek profile in both fast default, in blue and quiet in orange modes.
If we zoom in to the beginning of the disk, we see that the seek time of short seeks below 2. This makes sense: Short seeks are quiet even at full power, and adjacent-track seeks cannot be allowed to get slower because if an adjacent-track seek takes longer than the track skew, sequential throughput becomes halved. Finding track boundaries means finding the sector number of the first sector of every track. Designing an algorithm to find them reliably is surprisingly difficult. A track is a group of consecutive sectors that spans around one revolution of the disk, separated by track skew.
Throughput is proportional to track size if track skew is constant. The number of sectors per track is proportional to the sequential throughput of the disk. The drive spins at a constant angular speed, so sequential throughput varies with how much data is read per revolution.
Previous work has even used this relation in reverse to compute approximate track sizes by measuring sequential throughput [ 1 , 2 ]. The two figures above show a comparison between track sizes and a throughput benchmark for the same disk Toshiba MKGSX , showing that they look very similar.
Western Digital S25 GB. Track size monotonically decreases except for defects. Toshiba P 3TB. The region near track , is a region with many defective sectors causing the track size to drop.
All newer hard drives use zone bit recording, so earlier tracks have more sectors than later tracks. This sequence is usually not strictly monotonically decreasing because the number of sectors per track is often different on each recording surface, causing the track size to vary according to how tracks are arranged onto the surfaces.
For the Toshiba P and X, the plot looks shaded because the track size switches very frequently between several values according to which surface a track is placed on.
The WD S25, however, uses different track sizes on each surface, but tracks are ordered by decreasing size, not by physical location.
The result is a monotonically-decreasing track size curve, but highly irregular serpentine sizes causing an irregular seek profile. We will discuss this further in the section on measuring track layouts.
In the above plots, the x-axis is in units of track number not sector number. If the track density were constant, then track number would be linearly related to the physical radius of the track. Also, if the linear bit density were constant across the disk, then the track size would be proportional to the length and also radius of the track.
If the above two assumptions were true, I would expect to see the track size vs. But almost all of the drives show a slight convex shape. In the usually middle tracks of the disk, the disk head is designed to be aligned with the track, but at the inner ID and outer OD diameter positions, the head is rotated relative to the track because the head is mounted on a rotating arm , which causes the achievable bit density to be lower at the ID and OD.
Cordle et al. In the simplest case, a track consists of a group of consecutive sectors, separated by a track skew. Thus, we can detect track boundaries by finding the location of the skew an unusually large change in angular position between two nearby sectors. We can also assume that track sizes tend to be fairly similar adjacent tracks differ in size by less than a few tens of percent.
Unfortunately, none of these assumptions actually hold consistently across all drives and all tracks. I needed to manually fix some of the incorrectly-detected track boundaries. My current implementation does not work on the STA, but fortunately it is so small that it is feasible to manually determine the track size 26 sectors per track without zone bit recording. Even for drives with non-zero track skew, there are tracks with different skew.
Most drives have unusual sometimes near-zero skew at serpentine or zone boundaries, and skew can be fairly random when a track boundary occurs along with track slipping defect management or in the middle of a block of defective sectors when using sector-slipping defect management.
I implemented two algorithms. The first is a O lg tracksize algorithm to find a track boundary. A track boundary contains a track skew that increases the angular distance between two logical sectors if the two sectors cross a track boundary. The algorithm can then repeatedly split the region into two and test which half contains the track boundary binary search until the exact location is found. There are several speed optimizations on top of this basic algorithm.
The algorithm also predicts the location of the next track boundary track size tends to be constant within a serpentine and zone , and if the prediction is correct which is the common case , the next track boundary can be confirmed without a binary search.
Also, for large regions, it splits the region into more than two sub-regions because we can test multiple regions in the same revolution. Because track sizes tend to stay constant for relatively large regions, it may be profitable to not just predict the location of the next track boundary, but to predict that multiple upcoming tracks are all the same size.
If the prediction is correct, we double the prediction for the next iteration, otherwise we halve the number of skipped tracks and try again.
If the prediction is not exactly correct even for one track ahead, it falls back to the O lg tracksize algorithm described above to handle the change in track size. To verify the prediction, I verify that track boundaries exist at both n-1 and n tracks ahead Gim and Won only checks n tracks ahead. This was necessary because when track size changes, the change in track size multiplied by the number of tracks the algorithm chose to skip can be exactly a multiple of the previous track size, leading to an incorrect result.
For example, if the current track size is sectors and the algorithm wants to check whether 64 tracks or sectors ahead is still within the same zone, it would incorrectly conclude that the zone continues for 64 sector tracks if the reality were that there were 80 sector tracks instead. The probability of this occurring is low, but with millions of tracks per disk, I routinely saw a few incorrectly-predicted blocks for this reason.
Even doing two checks per prediction does not guarantee correctness. For example, if there were two consecutive tracks that were exactly half the size of its neighbours this is unlikely, but not impossible , both algorithms would incorrectly identify it as a single normal-sized track. In the most extreme cases namely, the blob of defects in one of my P drives , I had to resort to plotting an angular position plot of every sector in the region and identifying track boundaries manually.
Here are two examples where the track-finding algorithm fails. The first example illustrates difficulties when there are holes of missing sectors due to defects. The animation below shows the result of running the track boundary finding algorithm ten times on a region of the disk with one large hole and some small holes scattered around. The colour scheme is chosen to highlight the first few sectors of each track in dark blue.
Every run produces a different result as the algorithm often mistakes the hole for track skew. The second example does not involve defects. On the Samsung SVD, the last track of each zone may not be a complete track.
In this example, track is in a zone with sectors per track, but the final track of the zone is only 6 sectors long. The next zone has sectors per track, which is exactly 6 fewer sectors than the previous zone. An algorithm that looks for track boundaries would find one at precisely the expected location sectors past the end of track , not realizing that it missed a track boundary in between.
This case was found by manually inspecting all of the zone boundaries because I noticed very small tracks near some of the zone boundaries.
Now that we know the track boundaries and thus, the number of tracks , it would be interesting to try to estimate the track pitch, which is usually measured in tracks per inch TPI. Calculating the track pitch requires knowing the physical distance between the outer and inner tracks, which cannot be measured by timing alone.
This, I use photographs of the hard drives to estimate the physical size of the platters. This table shows the approximate measurements. Using these measurements, we can calculate the average track pitch track pitch can vary between the outer, middle, and inner tracks. It also says linear density is kbpi max. I do not know the reason for the difference, but possible causes include the manual publishing the maximum density of the densest region of the disk while I compute the average, or that I have over-estimated the physical area of the platter that is actually used for data for example, if the platter were capable of more storage than what was productized.
I can also use the same data to compute an estimate of the linear bit density for a particular track. Since we know the number of sectors per track, we can calculate the bit density of the outer track using the estimate for the outer diameter.
Some possible explanations include the outer track not having the highest linear density, or something to do with average density vs. After all this work, we finally have enough information to determine the track layout and number of recording surfaces and platters of the disks. The simple question of how many platters a hard drive contains turned out to be surprisingly hard to measure.
The graphs below plot both the track size and seek profile charts with a shared x-axis. The graphs show only the first few thousand tracks. The track size plot shows changes in the track size that indicate when head or zone changes occur, because a head change can also change the track size if the other recording surfaces has a different track size.
Changes in track skew not plotted here can also be used to locate head changes, which is especially useful for drives that use the same track size on all surfaces.
The seek profile gives an approximation of the radial location of the track. Tracks located nearest the outer edge have lower seek times. This is the same method used by Gim and Won to find track layouts [ 1 ]. Samsung HDSJ: 4 recording surfaces. This disk switches surfaces infrequently note the x-axis scale , and head switch coincides with zone boundaries. Western Digital S 3 recording surfaces.
Each surfaces is shown in a different colour in the seek profile. It shows four recording surfaces, each of which has a different track size. Starting at the beginning of the disk, tracks are filled on one surface from outside towards the inside seek time increases from track 0 to track , then moves to the outer track of the next surface. This pattern repeats four times indicating four surfaces before moving further inwards.
The next group of four zones starts at track Looking at the track size upper subplot, we can see that the track layout cycles through the four surfaces in the same order: The third surface has the highest density.
Regions of the same size are more likely to be on the same surface and adjacent than zones that differ in track size. Also, all serpentines have tracks that are ordered from outside to inside.
The Toshiba X 5TB drive has 10 recording surfaces, each having different track density. The track layout is different than the HDSJ.
0コメント