Is
It Memory or an I/O Device?
There are two
ways of viewing a DASD, either as an extension of memory or as an
Input / Output device. Both views are equally
valid; the one to use depends on the feature that is to be discussed. As a part of a memory system, the disk plays
a key role in virtual memory. However,
it is accessed as an I/O device, and its interfacing must be discussed in those
terms. This textbook will discuss disk
drives a number of times.
Early
History of the Disk and Disk Drive
As noted in
chapter 1 of this textbook, data storage reached a crisis stage in the early
1950’s, when the amount of physical floor space required to
house the punched cards holding the data became excessive. The only option was to develop new
technology. IBM addressed the problem
and delivered three new technologies: magnetic tape, the magnetic drum, and the
magnetic disk. Each of these were smaller and much faster to access than paper cards.
The first
magnetic tape drive was the IBM 729, released in 1952. While the dominant form of secondary data
storage for quite some time, magnetic tape drives are not much used any more. For that reason, we shall not discuss
magnetic tapes in any more detail.
The first
magnetic drum memory was the ERA 110, built for the U.S. Navy by Engineering
Research Associates of Minneapolis, MN.
The ERA 110 could store 1 million bits of information, approximately 125
kilobytes. One of the test engineers
present at the first sea trials of this device communicated the following to
the author of this textbook.
“The device spun
at a high rate in order to allow fast access to the data. At first, the trials were a success. When the ship turned rapidly, the gyroscope
effect took hold and ripped the drive out of its mountings.”
Magnetic drum
memory was announced by IBM in 1953, and first shipped in 1954. While the drum memory seemed to be a viable
competitor for the magnetic disk, its high cost and low data density limited
its use. The only significant use by IBM
for the drum memory was as the main memory for the IBM 650, a first–generation
computer announced in July 1953 and last manufactured in 1962.
The RAMAC
The origin of
the magnetic disk drive was a research project started in 1952 by a group from
IBM in San Jose, CA. One of its main
goals was to develop a better technology for data storage. Magnetic drums were considered, but the
choice was finally made to use a flat platter design, as first reported by
Jacob Rainbow of the National Bureau of Standards (now known as NIST) in
1952. The disk was correctly assessed as
having a bright future.
The first
commercial disk drive, called the RAMAC (Random–Access Method of Accounting Control) 305 was demonstrated on
September 13, 1956. The unit comprised
fifty aluminum disks of 24 inch diameter, coated on both sides with a magnetic
oxide material that was a variant of the paint used on the Golden Gate Bridge.
The RAMAC 305
took up the better part of a room and could store all of 5MB of data -- the
equivalent of 64,000 punch cards or 2,000 pages of text with 2,500 characters
per page. The drive system had an input/output data rate of roughly 10
kilobytes per second. It sold for about
$200,000 -- or you could lease it for about $3,200 a month. At that time, a good new car from General
Motors sold for about $2,000.
Here is an early
picture of the RAMAC, taken from the IBM archives [R100]. It is the same type of unit as is pictured in
chapter 1 of this textbook.
Figure: The
Disks of the RAMAC 305
Later disk designs would all have at least one read/write
head for each recording surface, or a pair of such heads for each platter. The problem to be solved before this
innovation took hold was how to suspend the heads over the recording
surface. The RAMAC took a very early
approach, suggested probably by the music “Juke Boxes” of the time. It has one set of two read/write heads that
served the entire drive.
Figure: The RAMAC Read/Write Heads [R100]
The IBM 1301
The RAMAC 305
was the precursor to the IBM 1301 disk storage unit. When released in 1961, the
1301 was the first storage system that used "flying heads" on
actuator arms to read and write data to its 50 24-inch magnetic platters. The
1301's head and actuator arm assembly looked something like a bread-slicing
machine turned on its side because each drive platter had its own read/write
head.
The 1301 had 13
times the capacity of the RAMAC, and its platters rotated at 1,800 rpm --
compared with a spindle speed of 100 rpm for the RAMAC -- allowing heads to
access the data more quickly.
Only two years
after creating the 1301, IBM built the first removable hard drive, the 1311.
The drive system, which shrunk storage technology from the size of a
refrigerator to the size of a washing machine, had six 14-inch platters and
contained a removable disk pack that had a maximum capacity of 2.6MB of data.
The 1311 remained in use through the mid-1970s.
The concept of
integrating the disks and the head arm assembly as a sealed unit was introduced
by IBM with the “Winchester Drive” in 1973.
Formally named the IBM 3340, this drive had two spindles, each of 30 MB
capacity; hence “30/30 Winchester”.
Though a project name, the term “Winchester” has been applied to the
design concept. Almost all modern disk
drives use the “Winchester design”.
While earlier designs provided for removable disks, these allowed the
disk to become dirty when exposed to the outside air. In the Winchester design, the entire assembly
can be removed, thus keeping the disks in a more protected environment. With this higher quality operating
environment, the read/write heads could be positioned closer to the disk
platters, allowing for higher recording densities.
The relationship
between head flying height and track density can be seen in the rough figure
below. Each head will react to tracks
within a specific angular dimension; think of a search light. The closer the head to the recording surface,
the smaller the spot containing elements to which the head will react. The head must not be able to react to more
than one data track at a time, hence its height
dictates the minimum track spacing.
More Developments
In 1979, Al Shugart, who had helped develop the RAMAC with IBM,
launched Seagate Technology Corp., which became the largest disk drive manufacturer
in the world. Soon thereafter, the
innovation floodgates opened. The "small form-factor" hard drive was
invented in 1980 by Seagate. That five–inch
ST506 drive held the same capacity as the RAMAC (5MB). From this point on, the mass market associated
with the newly arrived PC (Personal Computer) movement guaranteed rapid
evolution of the hard disk drive.
History
of Disk Costs
Here is a walk
through history [R101]. From the days of $10,000/MB to 2004 (from 2004 to
2009 the cost/GB has literally gone down to negligible levels), take a
look:
|
YEAR |
MANUFACTURER |
COST/GB |
|
1956 |
IBM |
$10,00,000 |
|
1980 |
North Star |
$193,000 |
|
1981 |
Morrow Designs |
$138,000 |
|
1982 |
Xebec |
$260,000 |
|
1983 |
Davong |
$119,000 |
|
1984 |
Pegasus (Great Lakes) |
$80,000 |
|
1985 |
First Class Peripherals |
$71,000 |
|
1987 |
Iomega |
$45,000 |
|
1988 |
IBM |
$16,000 |
|
1989 |
Western Digital |
$36,000 |
|
1990 |
First Class Peripherals |
$12,000 |
|
1991 |
WD |
$9,000 |
|
1992 |
Iomega |
$7,000 |
|
1994 |
Iomega |
$2000 |
|
1995 |
Seagate |
$850 |
|
1996 |
Maxtor |
$259 |
|
1997 |
Maxtor |
$93 |
|
1998 |
Quantum |
$43 |
|
1999 |
Fujitsu IDE |
$16 |
|
2000 |
Maxtor 7200rpm UDMA/66 |
$9.58 |
|
2001 |
Maxtor 5400 rpm IDE |
$4.57 |
|
2002 |
Western Digital 7200 rpm |
$2.68 |
|
2003 |
Maxtor 7200 rpm IDE |
$1.39 |
|
2004 |
Western Digital Caviar SE 7200rpm |
$1.15 |
More
recently disk prices have dropped even more.
In November 2007, the author of this textbook purchased an external USB
500 GB disk drive for about $500.00 (or $0.50 per gigabyte). On July 7, 2011, the web site for Office
Depot had the following prices:
a 500 GB USB external drive for $69.99 and a 1 TB USB external drive for
$119.99.
These prices are about five to seven cents per gigabyte.
We now
look at internal disk drives supporting the SATA standard. Also on July 7, 2011, these prices were
found. The WD Caviar Blue 500 GB drive
with a 16 MB buffer will sell for $38.00 and the 3 TB WD Caviar Green with a 64
MB buffer will sell for $140.
Do the math. This is very inexpensive.
How long
shall we have to wait for the first 1 petabyte drive?
Structure
of a Large Disk Drive
The typical large–capacity (and physically small) disk drive
has a number of glass platters with magnetic coating. These spin at a high rate varying between
3,600 rpm (60/second) and 15,000 rpm (250/second), with 7,200 rpm (120/second)
being average. This drawing shows a disk
with three platters and six surfaces. In
general, a disk drive with N platters will have 2·N surfaces, the top and bottom of each platter.
On early disk drives, before the introduction of sealed
drives, the top and bottom surfaces would not be used because they would become
dirty when the disk pack was removed.
The introduction of the Winchester design, with its sealed disks,
changed that for the better.
More
on Disk Drive Structure
Each surface is divided into a number of concentric tracks. Each track has a number of sectors. This figure shows an older style layout, in
which all tracks had the same number of sectors. In such a design, the time for a sector to
move under the read/write head does not depend on the track location,
simplifying the drive control logic.
A sector usually contains 512 bytes of data, along with a
header and trailer part. Modern designs
have moved towards larger sizes for sectors (say, 4,096) bytes, but the
traditional design has considerable history.
Modern disk designs have divided the surface into a number
of zones. This design is often called ZBR (Zoned Bit Recording). Within each zone, the sector count per track
is constant. However, the zones near the
outer rim of the disk have more sectors than the inner zones. This increases the linear density of the
sectors for the outer zones, keeping that density nearer that found in the
innermost zone, and making better use of the disk area. Here is a picture of a disk surface with
color added to highlight the five zones.
Note that the number of sectors per track increases as we move from the
inner zone to the outer zone.
Figure: Disk Surface with Five Zones
One downside to the ZBR strategy is that the complexity of
the disk–side controller must be increased.
In the older strategy, which might be called one–zone, the sectors moved
under the read/write heads at the same rate, without regard to track
number. In the above design, the rate is
constant within each zone, but the controller must allow for five distinct
rates. Given the ability to fabricate
complex electronics, this is deemed a fair trade–off.
The
Disk Surface
The standard definition of a hard disk drive is quite
simple.
“The recording medium for hard disk drives is basically a
very thin layer of magnetically hard material on a rigid substrate.” [R98]
Up to the mid 1990’s, aluminum was used exclusively for the
substrate. Recently, glass has attracted
interest for a substrate, mostly for its ability to be polished to a very fine
surface finish. Remember that the design
goal is to reduce the flying height of the read/write heads, and that requires
a smoother surface over which to fly.
Since the earliest designs, each recording surface of a disk
has been divided into concentric circles, called tracks. Optical disks, such as CDs and DVDs, as well
as vinyl records, have a single spiral track on each recording surface. According to Jacob [R99]
“The major reasons for this are partly historical … In the
early days of disk drives, when open–loop mechanical systems were used to
position the head, using concentric tracks seemed a
natural choice. Today, the closed–loop
embedded servo system … makes spiral track formatting feasible. Nonetheless concentric tracks are a
well–understood and well–used technology and seem easier to implement than
spiral tracks.”
Timing
of a Disk Transfer
Disks are electromechanical devices. The time for a data transfer has a number of
components, including time to decode the command, time to move the read/write
heads to the proper location on the disk, time to transfer the data to the
on–disk data buffers, and time to transfer the data across the interface to the
host–side interface. Here, we consider
two of the more important disk timings: seek time and rotational latency.
The idea of seek time
reflects the fact that, in order to access a disk track for either a read or
write operation, the heads must be moved to the track. This is a mechanical action, as the
read/write heads are physical devices.
Some early designs seem to have avoided this problem by having one
read/write head per track. This is not
an option for modern drives, which have thousands of tracks per surface.
There are two seek times typically quoted for a disk.
Track–to–track: the time to move the heads to the
next track over
Average: the average time to move
the heads to any track.
The rotational delay is due to the fact that the disk is spinning
at a fixed high speed. It takes a
certain time for a specific sector to rotate under the read/write heads. Suppose a disk rotating at 12,000 RPM. That is 200 revolutions per second. Each sector moves under the read/write heads
200 times a second, once every 0.005 second or every 5 milliseconds. The rotational
latency, or average rotational delay, is one half of the time for a
complete revolution of the disk. Here it
would be 2.50 milliseconds.
Rotational latency is a major component of the time to
access data on a disk. For that reason,
methods to minimize the rotational latency of a disk are under active
investigation. One way would be to speed
up the rotation of the disk.
Unfortunately, this leads to stress on the physical platters that become
unacceptable. This is seen in a design
choice for the 5¼ floppy drives, which were spun at a leisurely 360 rpm. The reason for this rate is that a
significantly higher rate would cause the disks to be torn apart by centrifugal
forces.
Another method to reduce rotational latency would be to add
a second actuator with its own set of read write heads to the disk drive. This would be positioned opposite the first
set.
Figure: Two Views of Dual Actuators
Such a design would immediately halve the rotational
latency, as the average time for a sector to move under one of the read/write
heads would be one quarter of a revolution.
For a 12,000 rpm disk, the rotational delay would be cut from 2.50 to
1.25 milliseconds. Put another way, the
reduction is from 2, 500, 000 nanoseconds to 1, 250, 000 nanoseconds. However, the addition of a second set of
heads raises the price by about 35%, not acceptable in the highly competitive
modern market. Conner Peripherals
introduced such a product in 1994, but it was not popular. The design has never been attempted since.
Tracks
and Cylinders
The idea of a cylinder is an artifact of the mechanical
nature of the actuators used to move the read/write heads on a modern
disk. It is faster to read sectors from
the same track than it is to move to another track, even one close by. We now consider the fact that standard disk
drives with concentric tracks and multiple recording surfaces have the same
number of tracks and track geometry on each of the surfaces.
When the actuator is functioning
properly, as it almost always is, each read/write head is over (or near to) the
same numbered track on its disk. This leads to the idea of a cylinder as the set of tracks that can
be read without significant repositioning of the read/write heads. In early disk designs, with wider tracks,
there would be no positioning required to move from surface to surface on a
cylinder. All that would be required was
an electronic switching of the active head, a matter of nanoseconds and not milliseconds.
We close this
part of the chapter with some related questions. How many bytes of data can be read from a
disk before the read/write heads must be moved?
What is the maximum data transfer rate from a multi–platter disk? How long can this rate be
sustained?
Disk Capacity
The first
question is how to calculate the capacity of a disk. Here are a number of equivalent ways,
assuming the standard 512 bytes per sector.
Disk Capacity = (number of surfaces)·(bytes per surface)
= (number of surfaces)·(tracks per surface)·(bytes per track)
= (number of surfaces)·(tracks per surface)· (sectors per track)·512
Here are data
from an earlier disk drive (now rather small).
8 surfaces
3196 tracks per
surface
132 sectors per
track
512 bytes per
sector
Surface
capacity = 3196·132·512 = 421872·512 = 210936·1024
bytes
= 210936 KB »
206.0 MB
Disk
capacity = 8·210936 KB = 1,687,488 KB » 1.61 GB
Computing Disk Maximum Transfer Rate
We now compute
the maximum transfer rate. This is
computed from the size of a track in bytes and the rotation rate of the
disk. For modern disks, the transfer rate
would depend on the zone in which the track is located, as each zone has a
different number of sectors per track.
For this example, we use the older disk with only one track size.
Disk rotation
rates are given in RPM (Revolutions per Minute). Common values are 3,600 RPM, 7,200 RPM, and
higher. 3,600 RPM is 60 revolutions per
second. 7,200 PRM is 120 per second. Consider our sample disk. Suppose it rotates at 7,200 RPM, which is one
revolution every (1/120) second.
One track contains 132·512 bytes = 66·1024 bytes = 66KB.
This track can be read in
(1/120) of a second.
The maximum data rate is 66 KB in (1/120) of a second.
120 ·
66 KB in one second.
7,920
KB per second = 7.73 MB per second.
Sustaining the
Maximum Data Transfer Rate
Recall the
definition of a cylinder as a set of tracks, one per surface. Each track can be read in one revolution of
the disk drive. The data rate can be
sustained for as long as it takes to read all tracks from the cylinder.
The number of
tracks per cylinder is always exactly the same as the number of surfaces in the
disk drive. Our sample drive has 8
surfaces; each cylinder has 8 tracks.
In
our sample drive, rotating at 7200 RPM or 120 per second:
Each track can be
read in 1/120 second.
The cylinder,
containing 8 tracks, is read in 8/120 second or 1/15 second.
The maximum data
transfer rate can be sustained for 1/15 second.
How Much Data
Can Be Transferred at This Rate?
Each track can
be read in 1/120 of a second. The
cylinder can be read in 1/15 second.
The cylinder contains 8 tracks of 66·1024
bytes, thus 8·66·1024 = 528 KB.
The Drive Cache
Modern disk
drives have a cache memory associated with the on–disk controller. Some manufacturers call this a “databurst
cache”. On modern drives, this can vary
between
2 MB and 64 MB; DRAM is quite cheap these days.
The disk–side
cache grew out of the small disk–side data buffers required to match the speed
of the disk internal transfers with the speed of the bus connecting it to the
controller logic on the host side.
Buffering is also required in those cases when the interface to the host
is busy at the time the disk read/write heads are ready to initiate a
transfer. On a disk write, the host
controller would transfer the data to the disk–side buffer for later writing to
the disk itself.
As DRAM prices
dropped, this simple data buffer grew into a full–fledged cache. The utility of this cache does not depend on
a high hit rate, as it does for the cache between primary memory and the CPU. One of its main uses is in allowing the disk
controller to prefetch data; that is, to copy the data into the buffer before
it is requested. This increases the
efficiency of disk reads. As a write
buffer, the cache speeds up disk writes.
We now have
quite a few levels of cache memory between the disk surfaces and the CPU. Here is a diagram showing the configuration
on a computer built in 2007. The disk
has a
16 MB DRAM cache. Each memory chip has a
4 Kb (kilobit) SRAM buffer. There are
two levels of cache between the main memory and the CPU.
The
Disk Directory Structure
We now turn our
attention to the disk directory structure.
Rather than discussing the more modern structures, such as NTFS, we
shall limit ourselves to older and simpler structures.
The
FAT (File Allocation Table)
Physically the
disk contains a large number of sectors, each of which contains either data or
programs. Logically the disk contains a
large number of files, also program and data.
The disk will have one or more index structures that associate files
with their sectors.
The two important structures are
the disk directory associates a file with its
first sector
the File Allocation Table maintains the “linked list” of sectors
for that file.
This example shows
the structure of the FAT.
Here
the disk directory indicates that the first sector for a file is at address
121.
The
FAT entry at 121 indicates that the next sector is at address 124.
The
FAT entry at 126 indicates that the next sector is at address 122.
The FAT entry at
122 indicates that sector 122 is the last for this file. Sector 125 is bad.
The FAT–16
system was implemented by Microsoft for early versions of MS–DOS. This system used a 16–bit index into the FAT. As there is one FAT entry per sector, this
makes this limits the sector count to 216. The maximum disk size is thus 216· 512 = 216· 29 = 225 =
25· 220
= 32 MB. In 1987, my brand–new PC/XT had
a 20MB disk! FAT–16 worked very well. What about a 40 MB disk? How about a 256 MB disk?
Few people in
the late 1980’s contemplated disk drives with capacities of over 100 GB,
but it was obvious that unmodified FAT–16 would not do the job. We consider two short–term remedies to this
problem, one transient and one with longer term consequences.
The first
solution was to partition a larger physical disk drive into two logical disk
drives.
A 40 MB disk
drive would support two logical disks, each with its own directory structure
and File Allocation Table. Using the
drive names suggested at the time, this 40 MB physical disk would support two
logical disks: Drive C with capacity of 32 MB, and
Drive D with capacity of 8 MB.
As a short–term
fix, this worked well. However, it just
raised the limit to 64 MB. This solution
was obsolete by about 1992. The main
problem with the FAT–16 system arose from the fact that each sector was
individually addressable. With 216
addresses available, we have a maximum of 216 sectors or 32 MB for
each logical drive.
The second
solution was to remove the restriction that each sector be addressable. Sectors were grouped into clusters, and only
clusters could be addressed. The number
of sectors a cluster contained was constrained to be a power of 2; so we had 2,
4, 8, 16, etc. sectors per cluster. The
effect on disk size is easy to see.
|
Sectors in Cluster |
1 |
2 |
4 |
8 |
16 |
32 |
64 |
|
Bytes in Cluster |
512 |
1024 |
2048 |
4096 |
8192 |
16384 |
32768 |
|
Disk Size |
32 MB |
64 MB |
128 MB |
256 MB |
512 MB |
1 GB |
2 GB |
In the early 1990’s, it seemed that this solution
would work for a while. Nobody back then
envisioned multi–gigabyte disk drives on personal computers. There is a problem associated with large
clusters; it is called “internal
fragmentation”.
This problem
arises from the fact that files must occupy an integer number of clusters.
Consider a data
file having exactly 6,000 bytes of data, with several cluster sizes.
|
Sectors in Cluster |
1 |
2 |
4 |
8 |
16 |
32 |
64 |
|
Bytes in Cluster |
512 |
1024 |
2048 |
4096 |
8192 |
16384 |
32768 |
|
Clusters needed |
12 |
6 |
3 |
2 |
1 |
1 |
1 |
|
File size on disk |
6144 |
6144 |
6144 |
8192 |
8192 |
16384 |
32768 |
|
Disk efficiency |
97.7% |
97.7% |
97.7% |
73.2% |
73.2% |
36.6% |
18.3% |
Security Issues: Erasing and Reformatting
Remember the
disk directory and the FAT. What happens
when a file is erased from the disk? The
data are not removed. Here is what
actually happens.
1. The file name is removed from the disk directory.
2. The FAT is modified to place all the sectors (clusters) for
that file
into a special
file, called the “Free List”.
3. Sectors do not have their data erased or changed until they
are allocated
to another file
and that a program writes data to that file.
For this reason, many companies sell utilities to
“Wipe the File” or “Shred the File”.
How about reformatting the disk? Certainly, that removes data. Again the data are not removed. The directory structure and FAT are
reinitialized and the free list reorganized.
The sectors containing the data are not overwritten until they are reused.