Future of Disk Drives

Future of Disk Drives

Szymon Jankowski

Abstract

In the past 50 years, disk drives have undergone great development in order to meet performance needs. Due to current design limitations, such asincreasing RPM while sacrificing reliability, traditional drive growth has been slowing. New storage optionsare being looked at as replacements to traditional disks. Hybrid disk combines the traditional disk with flash memory to act as a secondary cache, but this type of disk would only be a temporary solution. Solid State Disk has been catching up in power consumption and performance to traditional disks, but their lower capacity and higher cost deter most users from switching over. Flash memory has been developing rapidly in recent years and is the most likely candidate to succeed traditional disks. Unless traditional disks are redesigned to keep up with memory growth, it will be replaced within the next years with faster, more reliable alternatives.

Introduction

Traditional hard disk drives (HDDs) have become the most important persistent storage devices in modern computing due to their large capacity, high performance and reliability. As the performance gap between memory and disk drives widens to 6 orders of magnitude, a gap that continues to widen 50% per year, the future of disk drives requires examination. The current design of HDDs leave few areas open for improvement. Reducing disk access time, increasing platter RPM, and increasing areal densitywere considered for performance solutions, but each is accompanied by limitations. Since these improvements cannot be made to disk drives, new architectures and alternative storage mediums need to be considered. HDDs with multiple spindles or multiple actuators are the two main architectures under research. While not a new architecture, the inclusion of an RPM management system into the existing design is also under consideration. Among the new storage medium in research and development, there is magnetic random access memory, memristors, and phase change memory. Many of these new mediums types have the potential to one day replace disk drives, but they are all years away from practical application. The two frontrunners to replace HDDs are hybrid and solid state disk. These two disk drives, along with the other mediums and architecture changes, will be discussed after current HDD limitations are explained.

Figure 1: Disk drive architecture. [2]

Current Limitations

Figure 1 shows traditional disk drive architecture. The main components are platters, a spindle, a disk arm, disk heads, and a motor. The spindle rotates the platters at a constant RPM while the disk arm is powered by the motor. The motor moves the arm to move the disk heads to the platter. The disk heads are then moved by a Voice-Coil Motor (VCM). The VCM moves the disk heads over the recording surface of the platter to seek a target track.

Disk Access Time

Disk access time (Taccess) is composed of seek time (Tseek), rotational latency (Trotate), and data transfer time (Ttransfer) [2]. The seek time is the time it takes for the disk head to move to a specified track. When the head arrives at the target track, the time required for the platter to rotate the required sector underneath the head is the rotational latency. The data transfer time is the amount of data being transferred divided by the data transfer rate. The access time is expressed as follows.

Taccess = Tseek + Trotate + Ttransfer

Subsequently, the seek time can be further broken down into an acceleration time (tacc), a coast period (tcoast), a deceleration time (tdec), and a head settling time (tsettle). At the beginning of a track seek the disk head must first be accelerated. VCMs have reached a maximum head acceleration rate in order prevent further heat generation. The coast period is the point at which the VCM cuts power so the head can coast to the target rack. It is linear in the seek time. If a seek time is shorter than the threshold required to accelerate the head to the nominal maximal velocity, then the coast time can be eliminated from the equation. The acceleration phase would then be immediately followed by the deceleration phase. Deceleration time is equal to the acceleration time in order to slow the disk head from overshooting the target track. The settle time is the time it takes for the head to land on the correct track section. For random small requests, the seek time has remained the same due to the settle time largely remaining constant. Reducing settle time further does not seem possible. If the settle time cannot be reduced further, another possible way to decrease seek time is to decrease disk platter diameter. By decreasing platter diameter, all related seek times would decrease, with the exception of settle time. Unfortunately, reducing the platter diameter to smaller than 1.8 inches is difficult due to the size of associated mechanical components and heat dissipation issues.

Areal Density

Areal density (AD) in an HDD can be measured as the quantity of information bits on a length of track on a surface of the disk platter. The platter is usually made out of a magnetic metal. Each bit cell is composed of multiple magnetic grains. The AD is also affected by the number of grains in the surface, as well as the size of the read/write head. Current disk drives use longitudinal recording, which is estimated to reach between 100 to 200 gigabits per square inch. In 2005, Toshiba introduced the first prototype of a perpendicular recording hard drive. While the estimates for perpendicular storage capacity are constantly changing, they are predicted to reach a limit of 1 Terabits per square inch. The size of each grain cannot be scaled below a diameter of ten nanometers due to the superparamagnetic effect [4]. The superparamagnetic effect states that if there are too many grains in a cell, then the thermal energy produced by magnetization could be enough to overcome stored energy and reverse the magnetization, destroying the data stored there. To counter this negative effect, thestorage mediumneeds to have a larger magnetic region and the composing material a higher magnetic coercivity. Coercivity is the measure of resistance a material has to becoming demagnetized. There is a minimum size for a magnetic region at a given temperature and coercivity to resist demagnetization. If the region is any smaller, it is likely to be spontaneously demagnetized by local thermal fluctuations. Higher coercivity material allows the perpendicular write head to penetrate the medium more efficiently. These two factors combined are what make perpendicular recording possible. Perpendicular AD will eventually reach its limits, so the search for a future solution continues.

RPM and Dynamic Rotations per Minute

Increasing disk drive RPM would improve seek and write time. Modern hard drives have reached an average RPM of 10K, with some reaching as high as 15K. However, research has shown that disk drives rotating at speeds exceeding 20K RPM generate substantially more heat, have increased power consumption, noise, vibration, and other problems that attribute to lack of long-term reliability. The extra heat generated would factor into the superparamagnetic effect mentioned earlier. The increased power consumption would lead to a higher energy bill. Greater noise generation would become irritating. Vibration is a problem because the disk heads are integrated into a ceramic cylinder which includes an Air-Bearing Surface (ABS) facing the magnetic platter. The air trapped between the ABS and platter generates lift by taking advantage of the viscous properties of air being squeezed through the gap [2]. If the disk were to vibrate, it could compromise the very small gap between the head and platter, thus scratching the track and corrupting the data stored there.

Figure 2: Power state transition of disk drives. [2]

While not necessarily a solution to decrease data access time, the inclusion of an RPM manager in future hard drives would increase efficiency by lowering energy consumption and heat generation. By having dynamic rotations per minute (DRPM), a hard drive’s life span could be increased while decreasing energy consumption. Upon receiving a read/write request, a hard drive will enter an active power state in which the platters and disk arm are moved in order to fulfill the request. It is at this time the platters are rotating the fastest in order to find the required track. In the active state, the hard drive is also using the most energy and generating the most heat. After the request has been fulfilled, the drive enters an idle state in which the platters slow down (1). A slower RPM means less energy used and heat generated. In the event another request was to come in, it would require less energy to go from an idle state back to an active state (2). If no request comes in within a certain time threshold, then the idle drive would fully spin down into a standby state (3). The greatest energy consumption and heat generation comes from the transition from standby to active (4). If the standby state were to be eliminated, it would increase drive efficiency. Unfortunately, this solution is impractical for most desktop computers as they do not produce enough read/write requests daily to justify having the hard drive constantly working. Only server hard drives would benefit from DRPM since they are constantly spinning anyway.

HDD with Multiple Spindles

One of the newly proposed hard drive architectures is a multi spindle setup. This design would fit two smaller diameter platters into a standard size hard drive chassis. In order to keep the size of the chassis constant, the number of smaller platters is limited to two. By having smaller diameter platters, disk access time would improve since the tracks are shorter and require less time to move the disk head. Another use for the second platter would be to use it as the disk cache. While this would decrease drive storage capacity, by having a larger cache, it increases the probability of finding requested data in the cache. Accessing the cache takes less time than spinning up and searching the other platter, so that would be another improvement to data access time. This design is still in research and will require more time before it is applicable in computers.

Multiple Disk Actuators

Another disk drive architecture in research is the addition of a second actuator dedicated to reads. This read actuator would be fit into a standard size chassis and would share the same platter with the write actuator. The write actuator would always be positioned at the next available track, thus allowing for near-zero-access writes. This decreases disk access time by almost half since the write head would not have to seek for a free track. This design will also need more time to be researched before it is applicable.

New Storage Mediums

Magnetic platters are likely to reach their limitations soon, so numerous alternative storage mediums have been in research over the pastfew decades. A few of the most developed mediums are Magnetoresistive Random-Access Memory (MRAM), memristors, Phase Change Memory (PCM), and flash memory.

MRAM

MRAM drives are composed of two ferromagnetic plates separated by a thin insulating layer. One of these plates is set to a permanent magnetic polarity, while the polarity of the other plate can be changed. Data is stored on the basis of measuring the resistance between the two plates at a given cell by powering an associated transistor and passing a current through that cell. This method vastly differs from traditional disk. MRAM is nonvolatile, meaning its data is not susceptible to corruption. As long as the drive is not physically destroyed, the data could be stored indefinitely. As long as the drive is intact, it has near unlimited endurance. MRAM has faster data access times since there are no moving mechanical components involved, unlike traditional drives’ spinning platters and swinging disk arm. The design is also simplified without the mechanical components, leading to faster MRAM drive programming since there are less components to manage during read/write requests. Data access is random, as a transistor could be powered to search for the data at any location instead of having a disk head slowly work its way around a platter. MRAM storage density is currently limited by its cell size, which needs to reach 65nm in order to be put into mass production and compete with flash memory.

Memristors

Memristors, a portmanteau of “memory resistor”, were intended to be an alternate use for resistors, in which data would be stored based upon the resistance of the memristor. When an electrical current flows one direction through the memristor, the resistance would increase; when the current flows the opposite direction, the resistance would decrease. Even after powering off the device, the memristors would remember their resistance. This gives the data an almost unlimited lifespan, once again with the exception of physical destruction. Alloys and other potential memristor materials are being researched, so it will take a few years before a working prototype is produced.

Phase Change Memory

PCM is an exciting new proposed storage medium. It utilizes the large resistivity contrast between the crystalline (low) and amorphous (high) phases of a chalcogenide to stores bits in cells. A chalcogenide is the scientific term for an alloy along the GeTeSb2Te3 line, with the most commonly used alloy being Ge2Sb2Te3.

Figure 3: Graph showing chalcogenide phase change states. [3]

When a chalcogenide wafer is produced, all of the cells start in the crystalline state with low resistivity. To change the cell state from crystalline to amorphous, a low voltage current is applied until the cell reaches a voltage threshold (Vth), at which point the crystal melts (set). The now amorphous material then cools, giving that cell a higher resistivity. To switch the amorphous cell back to the crystalline state, a high voltage current is applied to melt the chalcogenide again until the crystal structure forms (reset). To read which state a cell is in, a very low voltage current is passed through the cell to measure resistivity (read).

Table 1: Comparative data between DRAM, NOR, NAND, and PCM. [3]

PCM has read time and data retention comparable to flash memory, but with lower energy consumption and potentially faster write time. The only downside to PCM currently is its cell sizes are about three times as large as that of flash memory. PCM needs a few more years of research in order to decrease cell size and increase storage capacity.

Flash Memory

Flash memory is probably the oldest and most well known alternative to traditional disk, with hybrid disk and Solid State Drives (SSDs) being released into the market in the 2000’s. NAND flash has faster erase/write times and higher data density than traditional HDDs. It is also nonvolatile, making NAND a better candidate for data storage. Flash memory does not have mechanical components, allowing the drives to be smaller and consume less power. This makes flash memory great for systems where size and power are of major importance, such as smart phones, digital cameras, etc.

Hybrid Disk

Hybrid disk is a combination of traditional disk with flash memory, where the flash acts as a second-level cache. The traditional disk portion operates the same as a regular HDD, but “hot” data items are stored in the flash memory. By having the flash cache, hit ratio is increased, reducing the number of read requests sent to the traditional disk’s platter. Traditional disk cache access time is expressed as follows.

Ttwo-layer = Hcache * Tcache + (1 - Hcache) * Taccess

The traditional disk cache hit ratio (Hcache) is multiplied by disk cache access time (Tcache), plus the disk access time (Taccess) multiplied by the probability the requested data is not stored in disk cache (1 - Hcache). When compared to hybrid disk cache access time:

Tthree-layer = Hcache * Tcache + (1 - Hcache) * (Hflash * Tflash + (1 – Hflash) * Taccess)

The flash hit ratio (Hcache) is multiplied by flash access time (Tcache), plus the disk access time multiplied by the probability the requested data is not in flash cache (1 – Hflash), which is then multiplied by the probability the requested data is not stored in disk cache. By having two caches, the probability that the requested data is on either cache is much higher, leading to potentially much faster access times. Using the manufacturer data for a Hitachi Ultrastar 15K147 hybrid drive in both equations, one can estimate a 2ms read/write time if treating the drive as a traditional disk; as a hybrid disk, .2ms read and .27ms write times. The hybrid disk’s access times are roughly 10 and 7.4 times faster, respectively.

Solid State Drives

Hybrid disk is merely a temporary solution for hard drive performance. Pure flash based disk drives, SSDs, do away with the traditional disk platter and use NAND flash to emulate a disk drive. Utilizing only flash memory for data storage does away with all the limitations and problems posed by traditional HDDs, with all the perks of flash: nonvolatile data storage, smaller physical size, lower power consumption, and better random read performance. To show just how much better SSD performance is in comparison to HDD, below is a graph comparing access times of several SSDs and HDDs.