Older analog read technologies are giving way to DSP-based digital read channels, which let you pack more data on your hard disk
Peter Wayner
Over the next several years, most hard drive manufacturers will abandon purely analog recording technology and begin shipping drives that use digital techniques. The new technology, called digital read channel, will at least double the amount of information that can be packed into each square inch of a hard drive platter, and there's no reason to doubt that further advances will produce even greater improvements in density. Such improvements in density are necessary as desktop machines move to applications that employ such rich data types as voice and video.
The key to the transition is a collection of low-cost drive controllers from companies such as Cirrus L
ogic (Fremont,CA), IBM Storage Systems Division (San Jose, CA), and Adaptec (Milpitas, CA) that incorporate digital signal processing to digitally process the signals from the disk fast enough to sustain the data rates needed by the main computer. These controllers use a digital signaling technique known as PRML (partial-response maximum-likelihood) that was originally used for communications with deep space probes, such as the Viking Lander. If the technique helped keep data clear through all the background interference from space, there is no reason why it can't help keep the data from disk drives clear, too.
PRML lets you pack more information into each track of a drive, because the algorithm allows the drive hardware to filter out the noise associated with densely packed data. The algorithms use knowledge of how nearby bits can blur together in order to clean up the signal and detect spurious signals. The new chips also pack more tracks on each platter by using digital techniques to align the drive
head over the correct track. When these two improvements are combined, they can lead to a doubling in the amount of data stored per disk. That means the cost of hard disk storage per byte will be cut in half.
Hard Drive Basics
To understand this approach, you need to begin with the physics of building a hard drive. A drive uses electromagnets to store data on the magnetically sensitive coatings of the platter. When the electromagnet is turned on, it generates a magnetic field. The molecules in the coating align themselves in the direction of the field in the same way that a compass turns to align itself with the magnetic field of the earth. Current can pass through the electromagnets in two directions, which means that resultant magnetic fields can have one of two opposite orientations. When you flip the polarity of the electromagnet, it will leave a similar pattern in the coating.
You read the data from the disk using the reverse of the write process. The magnetic fields in the coating gene
rate tiny electrical currents in the coils of the electromagnets as they pass under the read/write head. The presence or absence of an induced current determines the value of the spot.
At this level, both analog and digital hard drives and analog and digital tape drives operate in the same manner. The main difference is that the disk drives record bits as spots where the magnetic field saturates the coating and completely aligns all molecules. Old analog cassette decks stored the sound in the strength of the field that created larger and smaller magnetic patches. This technique is not generally used in disk drives because saturated blips are easier to read and write at high speed. This ease is also why digital tape players are making such inroads in the consumer audio world.
For years, drive manufacturers have continued to use tried-and-true analog read techniques while concentrating on other technologies to increase disk capacity. One major target has been reducing the distance between the head
and the platter. The closer the head, the smaller the magnetic spot and consequently, the more data that can be packed into one place. Of course, small spots also mean weaker induced fields when reading, so researchers have also concentrated on the development of more sensitive thin-film coatings. Research certainly continues in these two areas, but manufacturers are turning to PRML because the electronics necessary to do the more complicated digital calculations are finally becoming inexpensive enough to make digital drives a commercial possibility. Digital techniques let disk makers achieve significantly larger packing densities without investing the money in building physically more precise mechanisms or developing more sensitive recording media.
Encoding Data
The process of converting bits into magnetic marks on a disk involves at least three different layers of algorithms. On the highest level is the operating system, which writes the file (which may already have been compressed by software) t
o the disk electronics. When these bits are handed to the drive, the drive electronics then encodes them with ECCs (error-corrected codes), such as Reed-Solomon codes. ECCs have redundant bits that let the hardware reconstruct the original data when errors occur. The codes are a well-studied branch of mathematics. They often amount to constructing over-determined systems of equations where m equations and n unknowns exist, and m is greater than n.
This resulting bit stream is then converted into RLL codes, which add redundant bits to guarantee that not too many 1s or 0s fall in a row. The process for writing to the disk generates a new magnetic mark for every time a 1 occurs. If a 1 doesn't occur often enough, the clock that tries to count the position along the drive starts to drift, and the drive may drop a bit. Cirrus Logic uses an RLL in which every 1 is separated by at least one 0 and no more than seven 0s. During the course of this process, the bits increase in number by 50 percent. This loss of
packing density is necessary to guarantee the accuracy of the data.
The effects of writing bits on a disk can be seen in the figure "Sampling Signals." The curve represents the strength of the magnetic mark at a particular point on the track. Some of the peaks point up and others point down. Drive controllers usually alternate positive and negative peaks to help distinguish them.
The old generation of technology used two peak-detection circuits to determine whether a peak occurred. One circuit would check to see if the signal was above a preset level, and the second would take the derivative of the signal and look for a 0. This worked well if the peaks were spread apart, but as they grew closer, the mingling of two peaks would often add false peaks.
The Role of PRML
The main problem with crowded signals is that the read electronics do not see the smooth curves as shown in the figure "Sampling Signals." The heads can only provide a few samples for each peak. Thus, as you crowd more bits
into a smaller area, you need some method of determining to which peak a sample belongs. This is the role of the PRML algorithms in digital read channels. The algorithms determine whether each pair of samples belong to the same peak or to two different peaks. More important, the algorithms adjust themselves when errors occur, because it is not uncommon for noise to skew samples and add false peaks.
The PRML algorithms used in digital-read-channel controllers are similar to the algorithms that were developed in the sixties and used to process convolutional codes. These codes built in a certain amount of redundancy and error correction by writing each bit down as the sum of several of the previous bits. For instance, one simple convolutional code might write bit i as the sum of bit i, bit i-1, and bit i-3. Bit i can be recovered at read time by subtracting out the values of bit i-1 and bit i-3. Normally, these convolutional codes will store each bit as the result of two or more different polynomial equa
tions based on the previous n bits. The more complex the polynomials, the more resistant the code can be to error.
If there is no noise, these convolutional codes can be decoded by reversing the equation and subtracting out the values of the previous bits that have already been decoded. But if there is noise, the algorithm can use the extra information present to determine which bits were flipped. It tries to identify the most likely bit that caused the error. The algorithms for doing this are often known as Viterbi algorithms. In the simplest form, they amount to checking all possible errors and looking for the best match to the data in question. This technique is useful if you are using a small polynomial involving only a few bits that could go wrong. In the most complex cases, the algorithms perform like dynamic programming algorithms.
Convolutional codes are good models for disk drives where tightly packed data may create peaks that overlap. The tail of one bit can often interfere with the s
ignal of another. More important, each peak can contain several sample points. The algorithm's job is to determine when and how noise skews some of the sample points (see the text box "PRML at Work").
Embellishing PRML
Some manufacturers have developed proprietary approaches to these PRML algorithms. Cirrus touts its SofTarget feature that allows the electronics to tune itself on various drives. The electronics can use a range of different polynomials in the convolutional codes and also set the expected values for each peak. Each drive can tune itself and choose the best values. One drive might decide that its peaks normally generate + or - 3.7 for the first sample and + or - 3.4 for the second sample after it writes and reads a calibration pattern. Another might choose the pair + or - 3.2 and + or - 3.5.
Cirrus believes that the SofTarget technology gives drive manufacturers the ability to relax their tolerances on the magnetic media and the heads. The heads and the media do not have to perf
orm in as narrow a range as before.
The Electronics
Companies such as IBM, Cirrus, and Adaptec like to characterize digital-read-channel systems as using digital signal processing. While this is true, the use of these words can cause some confusion because the electronics built into the drive hardly resemble the DSP (digital signal processor) chips that are common today. The standard DSP is a stand-alone computer with some beefed up circuitry for doing simultaneous multiplication and addition.
On the other hand, these drive controllers can do only one thing--interpret the signals coming off the disk. The electronics are hard-wired circuits that can only be reprogrammed in small ways. The circuits must be simple because they have to process the data at high rates. The Cirrus CL-SH4400, for instance, can put out data at up to 64 Mbps at channel frequencies of up to 96 MHz.
The work that the CL-SH4400 does can be broken up into five basic steps. First, a companion chip, the VM6400 from VT
C (Bloomington, MN), converts the raw signal from analog into 6 bits of digital representation. When this arrives at the CL-SH4400, the signal is equalized and smoothed before the PRML algorithm attempts to filter out the errors. This is because the response of the head to the magnetic field is often not smooth. Finally, the chip removes the RLL codes and presents the result in ordinary form.
In addition, Cirrus makes the CL-SH3300, a one-chip version of the CL-SH4400/VM6400 pair that performs at a slower rate of 40 to 48 Mbps. A 64-Mbps derivative of the CL-SH3300 will be available soon, and Cirrus plans to push the two-chip version to even higher rates. Both packages are made in CMOS.
The Future
The emergence of PRML drives is likely to be responsible for major gains in price/performance of hard drives over the next several years. At this point, Cirrus claims that new drives can increase in capacity by up to 50 percent if the company uses the CL-SH4400 SofTarget technology with conventional
heads and magnetic media.
In the future, further gains will come when drive manufacturers can build magneto-resistive heads. These heads contain special resistors that change their resistance as they pass through the magnetic field caused by a magnetic mark on the disk platter. These heads are better tuned for PRML algorithms because they generate smoother signals. When these heads become common, drives should be able to pack twice as much information again into each platter.
The simple PRML algorithms used in this first round of drives is just the beginning. More complicated coding algorithms can result in greater densities of data. IBM's Almaden Research Center (San Jose, CA) is experimenting with a trellis-coded PRML system that uses a more sophisticated coding algorithm that is commonly used in high-speed modems. Its system uses a technique called Matched Spectral Nulls that more accurately tunes the coding algorithm to the read-head response. The algorithm knows which types of errors are m
ore likely to happen, and it can recover those better. IBM is reporting that experiments show that at least 15 percent more information can be packed into each track, and more important, off-track performance is improved, allowing narrower track widths.
These algorithms promise to reduce the amount of error correction that is built into the system. At this time, the three levels of codes--ECC, RLL, and PRML--significantly increase the number of actual bits that are written to the disk. If the coding process can be more properly tuned to the channel, then more data can be written.
The future drives will no doubt include even more innovations in coding technology. Digital disk electronics was not possible until recently; you couldn't fabricate chips with enough transistors to do the necessary computations. Now that this technology has arrived, there will be plenty of room in the future for new systems that incorporate better algorithms and more robust codes.
Read Technology
Analog
-- Is simple to implement
-- Needs clean samples
-- Can't handle closely packed bits
Digital
-- Requires fast processors
-- Can filter noise and other interference
-- Allows denser packing of bits
Figure: Sampling Signals
As laid down by the write circuitry, 1 bit on a hard drive corresponds to magnetized spots on the surface of the disk, while 0 bits corresponds to the absence of magnetization. The signal on the disk resembles a smooth curve. The read electronics, however, doesn't see a smooth curve. Rather, it samples the signal at discrete points. Interpreting these samples gets tricky as increasing bit densities pushes the peaks closer together.
Figure: Effects of Digital Read Channel
Digital technology lets disk makers pack bits more closely on individual tracks and also pack tracks more closely. The advent of magneto-resistive heads will further increase the packing density of hard disk platters.
Peter Wayn
er is a BYTE consulting editor. He can be reached on the Internet at
pcw@access.digex.net
or on BIX as "pwayner."
Digital Hard Drives
