Author Archives: kc7txm

About kc7txm

Matt Karls has a PhD in IT and is an Electrical Engineer. He works in management within the software development, IT and SEO fields and is the owner of Karls Technology. He has four kids and lives in the Phoenix metro area (when he is not travelling around to our different offices).

Hard Drive Error Correcting Code (ECC)

The basis of all error detection and correction in hard disks is the inclusion of redundant information and special hardware or software to use it. Each sector of data on the hard disk contains 512 bytes, or 4,096 bits, of user data. In addition to these bits, an additional number of bits are added to each sector for the implementation of error correcting code or ECC (sometimes also called error correction code or error correcting circuits). These bits do not contain data; rather, they contain information about the data that can be used to correct any problems encountered trying to access the real data bits.

There are several different types of error correcting codes that have been invented over the years, but the type commonly used on PCs is the Reed-Solomon algorithm, named for researchers Irving Reed and Gustave Solomon, who first discovered the general technique that the algorithm employs. Reed-Solomon codes are widely used for error detection and correction in various computing and communications media, including magnetic storage, optical storage, high-speed modems, and data transmission channels. They have been chosen because they are easier to decode than most other similar codes, can detect (and correct) large numbers of missing bits of data, and require the least number of extra ECC bits for a given number of data bits. Look in the memory section for much more general information on error detection and correction.

When a sector is written to the hard disk, the appropriate ECC codes are generated and stored in the bits reserved for them. When the sector is read back, the user data read, combined with the ECC bits, can tell the controller if any errors occurred during the read. Errors that can be corrected using the redundant information are corrected before passing the data to the rest of the system. The system can also tell when there is too much damage to the data to correct, and will issue an error notification in that event. The sophisticated firmware present in all modern drives uses ECC as part of its overall error management protocols. This is all done “on the fly” with no intervention from the user required, and no slowdown in performance even when errors are encountered and must be corrected.

The capability of a Reed Solomon ECC implementation is based on the number of additional ECC bits it includes. The more bits that are included for a given amount of data, the more errors that can be tolerated. There are multiple trade offs involved in deciding how many bits of ECC information to use. Including more bits per sector of data allows for more robust error detection and correction, but means fewer sectors can be put on each track, since more of the linear distance of the track is used up with non-data bits. On the other hand, if you make the system more capable of detecting and correcting errors, you make it possible to increase areal density or make other performance improvements, which could pay back the “investment” of extra ECC bits, and then some. Another complicating factor is that the more ECC bits included, the more processing power the controller must possess to process the Reed Solomon algorithm. The engineers who design hard disks take these various factors into account in deciding how many ECC bits to include for each sector.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

VESA Local Bus

The VESA Local Bus, also called VL-Bus or more commonly VLB, was the first local bus used on PCs. Introduced in 1992, VLB video became very popular during the heyday of the 486, in particular 1993 to 1994. VLB cards can be easily identified by their longer connectors, compared to standard ISA card slots. See here for details on the VESA local bus.

VLB video cards provide, in general, much better performance than ISA cards. This is primarily due to the fact that the 32-bit local bus used by VLB cards allows for several times more data throughput between the card and the processor than ISA allows. VLB has however had its own share of problems. In particular, VLB video cards may cause reliability problems in motherboards running at 40 or 50 MHz.

Many VLB cards are very good performers, but are hampered by their general age, along with that of the motherboards they run in; most are at least four years old and new development of better and faster chipsets is entirely in the PCI world now. Still, despite the fact that VLB is older than PCI, it can provide quite acceptable performance (although probably fewer features and less video memory). VLB is much closer to PCI than it is to ISA. Any system that will support VLB should be using it for the video card; the performance improvement over ISA is substantial in most cases.

Note: VESA Local Bus video is generally limited to 486 PCs (or other motherboards that use a fourth-generation processor). The vast majority of Pentiums and later PCs use PCI (or AGP) and do not support VLB at all, although there are some very old Pentium systems that are VLB-based.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

Hard Drive Spindle Speed

As hard disks become more advanced, virtually every component in them is required to do more and work harder, and the spindle motor is no exception. As discussed in detail here, increasing the speed at which the platters spin improves both positioning and transfer performance: the data can be read off the disk faster during sequential operations, and rotational latency–the time that the heads must wait for the correct sector number to come under the head–is also reduced, improving random operations. For this reason, there has been a push to increase the speed of the spindle motor, and more than at any other time in the past, hard disk spin speeds are changing rapidly.

At one time all PC hard disks spun at 3,600 RPM; in fact, for the first 10 years of the PC’s existence, that was all there was. One reason for this is that their designs were based on the old designs of large, pre-PC hard disks that used AC motors, and standard North American AC power is 60 Hz per second: 3,600 RPM. In the early 1990s manufacturers began to realize how much performance could be improved by increasing spindle speeds. The next step up from 3,600 RPM was 4,500 RPM; 5,400 RPM soon followed and became a standard for many years. From there speeds have steadily marched upwards. Usually, faster PC hard disk speeds “debut” on SCSI drives that are used in higher-performance applications, and then filter down to IDE/ATA a few years later. At one time 7,200 RPM spindles were only found on top-of-the-line SCSI drives; they are now being used in consumer IDE/ATA disks sold at retail while SCSI has moved on to loftier heights. This table shows the most common PC spindle speeds, their associated average rotational latency, and their typical applications as of early 2000:

Spindle Speed (RPM)Average Latency (Half Rotation) (ms)Typical Current Applications
3,6008.3Former standard, now obsolete
4,5006.7IBM Microdrive, laptops
5,4005.6Low-end  IDE/ATA, laptops
7,2004.2High-end IDE/ATA, Low-end SCSI
10,0003.0High-end SCSI
12,0002.5High-end SCSI
15,0002.0Top-of-the-line SCSI

Note: Hard disks for laptops and specialty applications come in a wide variety of spindle speeds, even beyond the several speeds listed above. I have not exhaustively researched and listed these here.

Increasing spindle motor speed creates many design challenges, particularly aimed at keeping vibration and heat under control. As discussed here, when the motor spins faster these become more of an issue; some high-end drives have very serious heat, vibration and noise problems that require special mounting and cooling work to allow them to run without problems. To some extent, there is a trade off between spindle speed, and the heat and noise issue. Engineers generally focus on keeping these matters under control, and usually improve them significantly after the first generation of drives at any given spindle speed. However, in some applications, using a slower and quieter drive can make sense.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

VLB for x86

VESA Local Bus (VLB)

The first local bus to gain popularity, the VESA local bus (also called VL-Bus or VLB for short) was introduced in 1992. VESA stands for the Video Electronics Standards Association, a standards group that was formed in the late eighties to address video-related issues in personal computers. Indeed, the major reason for the development of VLB was to improve video performance in PCs.

The VLB is a 32-bit bus which is in a way a direct extension of the 486 processor/memory bus. A VLB slot is a 16-bit ISA slot with third and fourth slot connectors added on the end. The VLB normally runs at 33 MHz, although higher speeds are possible on some systems. Since it is an extension of the ISA bus, an ISA card can be used in a VLB slot, although it makes sense to use the regular ISA slots first and leave the (small number of) VLB slots open for VLB cards, which won’t work in an ISA slot of course. Use of a VLB video card and I/O controller greatly increases system performance over an ISA-only system.

While VLB was extremely popular during the reign of the 486, with the introduction of the Pentium and its PCI local bus in 1994, wholesale abandonment of the VLB began in earnest. While Intel pushing PCI was one reason why this happened, there were also several key problems with the VLB implementation. First, the design was strongly based on the 486 processor, and adapting it to the Pentium caused a host of compatibility and other problems. Second, the bus itself was tricky electrically; for example, the number of cards that could be used on the bus was low (often only two or even one), and occasionally there could be timing problems on the bus when more than one card was used. Finally, the bus did not support bus mastering properly since there was no good arbitration scheme, and did not support Plug and Play.

Today VLB is obsolete for new systems; even the latest 486 motherboards use PCI, and all Pentiums and higher use PCI. However, these systems do still offer reasonable performance, and are now plentiful and very inexpensive–if you can still find them.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

RAID Levels 03 and 30

RAID Levels 0+3 (03 or 53) and 3+0 (30)

Common Name(s): The most confusing naming of any of the RAID levels. :^) In an ideal world, this level would be named RAID 0+3 (or 03) or RAID 3+0 (30). Instead, the number 53 is often used in place of 03 for reasons I have never been able to determine, and worse, 53 is often actually implemented as 30, not 03. As always, verify the details of the implementation to be sure of what you have.

Technique(s) Used: Byte striping with dedicated parity combined with block striping.

Description: RAID 03 and 30 (though often called 53 for a reason that utterly escapes me) combine byte striping, parity and block striping to create large arrays that are conceptually difficult to understand. :^) RAID 03 is formed by putting into a RAID 3 array a number of striped RAID 0 arrays; RAID 30 is more common and is formed by striping across a number of RAID 3 sub-arrays. The combination of parity, small-block striping and large-block striping makes analyzing the theoretical performance of this level difficult. In general, it provides performance better than RAID 3 due to the addition of RAID 0 striping, but closer to RAID 3 than RAID 0 in overall speed, especially on writes. RAID 30 provides better fault tolerance and rebuild performance than RAID 03, but both depend on the “width” of the RAID 3 dimension of the drive relative to the RAID 0 dimension: the more parity drives, the lower capacity and storage efficiency, but the greater the fault tolerance. See the examples below for more explanation of this.

Most of the characteristics of RAID 0+3 and 3+0 are similar to those of RAID 0+5 and 5+0. RAID 30 and 03 tend to be better for large files than RAID 50 and 05.

Controller Requirements: Generally requires a high-end hardware controller.

Hard Disk Requirements: Number of drives must be able to be factored into two integers, one of which must be 2 or higher and the other 3 or higher (you can make a RAID 30 array from 10 drives but not 11). Minimum number of drives is six, with the maximum set by the controller.

Array Capacity: For RAID 03: (Size of Smallest Drive) * (Number of Drives In Each RAID 0 Set) * (Number of RAID 0 Sets – 1). For RAID 30: (Size of Smallest Drive) * (Number of Drives In Each RAID 3 Set – 1) * (Number of RAID 3 Sets).

For example, the capacity of a RAID 03 array made of 15 18 GB drives arranged as three five-drive RAID 0 sets would be 18 GB * 5 * (3-1) = 180 GB. The capacity of a RAID 30 array made of 21 18 GB drives arranged as three seven-drive RAID 3 sets would be 18 GB * (7-1) * 3 = 324 GB. The same 21 drives arranged as seven three-drive RAID 3 sets would have a capacity of 18 GB * (3-1) * 7 = “only” 252 GB.

Storage Efficiency: For RAID 03: ( (Number of RAID 0 Sets – 1) / Number of RAID 0 Sets). For RAID 30: ( (Number of Drives In Each RAID 3 Set – 1) / Number of Drives In Each RAID 3 Set).

Taking the same examples as above, the 15-drive RAID 03 array would have a storage efficiency of (3-1)/3 = 67%. The first RAID 30 array, configured as three seven-drive RAID 3 sets, would have a storage efficiency of (7-1)/7 = 86%, while the other RAID 30 array would have a storage efficiency of, again, (3-1)/3 = 67%.

Fault Tolerance: Good to very good, depending on whether it is RAID 03 or 30, and the number of parity drives relative to the total number. RAID 30 will provide better fault tolerance than RAID 03.

Consider the two different 21-drive RAID 30 arrays mentioned above: the first one (three seven-drive RAID 3 sets) has higher capacity and storage efficiency, but can only tolerate three maximum potential drive failures; the one with lower capacity and storage efficiency (seven three-drive RAID 3 sets) can handle as many as seven , if they are in different RAID 3 sets. Of course few applications really require tolerance for seven independent drive failures! And of course, if those 21 drives were in a RAID 03 array instead, failure of a second drive after one had failed and taken down one of the RAID 0 sub-arrays would crash the entire array.

Availability: Very good to excellent.

Degradation and Rebuilding: Relatively little for RAID 30 (though more than RAID 10); can be more substantial for RAID 03.

Random Read Performance: Very good, assuming RAID 0 stripe size is reasonably large.

Random Write Performance: Fair.

Sequential Read Performance: Very good to excellent.

Sequential Write Performance: Good.

Cost: Relatively high due to requirements for a hardware controller and a large number of drives; storage efficiency is better than RAID 10 however and no worse than any other RAID levels that include redundancy.

Special Considerations: Complex and expensive to implement.

Recommended Uses: Not as widely used as many other RAID levels. Applications include data that requires the speed of RAID 0 with fault tolerance and high capacity, such as critical multimedia data and large database or file servers. Sometimes used instead of RAID 3 to increase capacity as well as performance.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

RAID Levels 01 and 10

RAID Levels 0+1 (01) and 1+0 (10)

Common Name(s): RAID 0+1, 01, 0/1, “mirrored stripes”, “mirror of stripes”; RAID 1+0, 10, 1/0, “striped mirrors”, “stripe of mirrors”. Labels are often used incorrectly; verify the details of the implementation if the distinction between 0+1 and 1+0 is important to you.

Technique(s) Used: Mirroring and striping without parity.

Description: The most popular of the multiple RAID levels, RAID 01 and 10 combine the best features of striping and mirroring to yield large arrays with high performance in most uses and superior fault tolerance. RAID 01 is a mirrored configuration of two striped sets; RAID 10 is a stripe across a number of mirrored sets. RAID 10 and 01 have been increasing dramatically in popularity as hard disks become cheaper and the four-drive minimum is legitimately seen as much less of an obstacle. RAID 10 provides better fault tolerance and rebuild performance than RAID 01. Both array types provide very good to excellent overall performance by combining the speed of RAID 0 with the redundancy of RAID 1 without requiring parity calculations.

RAID Levels 01 and 10

This illustration shows how files of different sizes are distributed between the drives on an eight-disk RAID 0+1 array using a 16 kiB stripe size for the RAID 0 portion. As with the RAID 0 illustration, the red file is 4 kiB in size; the blue is 20 kiB; the green is 100 kiB; and the magenta is 500 kiB, with each vertical pixel representing 1 kiB of space. The large, patterned rectangles represent the two RAID 0 “sub arrays”, which are mirrored using RAID 1 to create RAID 0+1.
The contents of the striped sets are thus identical. The diagram for RAID 1+0
would be the same except for the groupings: instead of two large boxes dividing the drives horizontally, there would be four large boxes dividing the drives vertically into mirrored pairs. These pairs would then be striped together to form level 1+0. Contrast this diagram to the ones for RAID 0 and RAID 1.

Controller Requirements: Almost all hardware controllers will support one or the other of RAID 10 or RAID 01, but often not both. Even low-end cards will support this multiple level, usually RAID 01. High-end cards may support both 01 and 10.

Hard Disk Requirements: An even number of hard disks with a minimum of four; maximum dependent on controller. All drives should be identical.

Array Capacity: (Size of Smallest Drive) * (Number of Drives ) / 2.

Storage Efficiency: If all drives are the same size, 50%.

Fault Tolerance: Very good for RAID 01; excellent for RAID 10.

Availability: Very good for RAID 01; excellent for RAID 10.

Degradation and Rebuilding: Relatively little for RAID 10; can be more substantial for RAID 01.

Random Read Performance: Very good to excellent.

Random Write Performance: Good to very good.

Sequential Read Performance: Very good to excellent.

Sequential Write Performance: Good to very good.

Cost: Relatively high due to large number of drives required and low storage efficiency (50%).

Special Considerations: Low storage efficiency limits potential array capacity.

Recommended Uses: Applications requiring both high performance and reliability and willing to sacrifice capacity to get them. This includes enterprise servers, moderate-sized database systems and the like at the high end, but also individuals using larger IDE/ATA hard disks on the low end. Often used in place of RAID 1 or RAID 5 by those requiring higher performance; may be used instead of RAID 1 for applications requiring more capacity.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

RAID Levels Comparison

Summary Comparison of RAID Levels

Below you will find a table that summarizes the key quantitative attributes of the various RAID levels for easy comparison. For the full details on any RAID level, see its own page, accessible here. For a description of the different characteristics, see the discussion of factors differentiating RAID levels. Also be sure to read the notes that follow the table:

Comparison of all RAID Levels

Notes on the table:

  • For the number of disks, the first few valid sizes are shown; you can figure out the rest from the examples given in most cases. Minimum size is the first number shown; maximum size is normally dictated by the controller. RAID 01/10 and RAID 15/51 must have an even number of drives, minimum 6. RAID 03/30 and 05/50 can only have sizes that are a product of integers, minimum 6.
  • For capacity and storage efficiency, “S” is the size of the smallest drive in the array, and “N” is the number of drives in the array. For the RAID 03 and 30, “N0” is the width of the RAID 0 dimension of the array, and “N3” is the width of the RAID 3 dimension. So a 12-disk RAID 30 array made by creating three 4-disk RAID 3 arrays and then striping them would have N3=4 and N0=3. The same applies for “N5” in the RAID 05/50 row.
  • Storage efficiency assumes all drives are of identical size. If this is not the case, the universal computation (array capacity divided by the sum of all drive sizes) must be used.
  • Performance rankings are approximations and to some extent, reflect my personal opinions. Please don’t over-emphasize a “half-star” difference between two scores!
  • Cost is relative and approximate, of course. In the real world it will depend on many factors; the dollar signs are just intended to provide some perspective.

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

Some Computer Science Issues in Ubiquitous Computing

by Mark Weiser
March 23, 1993

[to appear in CACM, July 1993]

Ubiquitous computing is the method of enhancing computer use by making many computers available throughout the physical environment, but making them effectively invisible to the user. Since we started this work at Xerox PARC in 1988, a number of researchers around the world have begun to work in the ubiquitous computing framework. This paper explains what is new and different about the computer science in ubiquitous computing. It starts with a brief overview of ubiquitous computing, and then elaborates through a series of examples drawn from various subdisciplines of computer science: hardware components (e.g. chips), network protocols, interaction substrates (e.g. software for screens and pens), applications, privacy, and computational methods. Ubiquitous computing offers a framework for new and exciting research across the spectrum of computer science.

A few places in the world have begun work on a possible next generation computing environment in which each person is continually interacting with hundreds of nearby wirelessly interconnected computers. The point is to achieve the most effective kind of technology, that which is essentially invisible to the user. To bring computers to this point while retaining their power will require radically new kinds of computers of all sizes and shapes to be available to each person. I call this future world “Ubiquitous Computing” (short form: “Ubicomp”) [Weiser 1991]. The research method for ubiquitous computing is standard experimental computer science: the construction of working prototypes of the necessary infrastructure in sufficient quantity to debug the viability of the systems in everyday use, using ourselves and a few colleagues as guinea pigs. This is an important step towards insuring that our infrastructure research is robust and scalable in the face of the details of the real world.

The idea of ubiquitous computing first arose from contemplating the place of today’s computer in actual activities of everyday life. In particular, anthropological studies of work life [Suchman 1985, Lave 1991] teach us that people primarily work in a world of shared situations and unexamined technological skills. However the computer today is isolated and isolating from the overall situation, and fails to get out of the way of the work. In other words, rather than being a tool through which we work, and so which disappears from our awareness, the computer too often remains the focus of attention. And this is true throughout the domain of personal computing as currently implemented and discussed for the future, whether one thinks of PC’s, palmtops, or dynabooks. The characterization of the future computer as the “intimate computer” [Kay 1991], or “rather like a human assistant” [Tesler 1991] makes this attention to the machine itself particularly apparent.

Getting the computer out of the way is not easy. This is not a graphical user interface (GUI) problem, but is a property of the whole context of usage of the machine and the affordances of its physical properties: the keyboard, the weight and desktop position of screens, and so on. The problem is not one of “interface”. For the same reason of context, this was not a multimedia problem, resulting from any particular deficiency in the ability to display certains kinds of realtime data or integrate them into applications. (Indeed, multimedia tries to grab attention, the opposite of the ubiquitous computing ideal of invisibility). The challenge is to create a new kind of relationship of people to computers, one in which the computer would have to take the lead in becoming vastly better at getting out of the way so people could just go about their lives.

In 1988, when I started PARC’s work on ubiquitous computing, virtual reality (VR) came the closest to enacting the principles we believed important. In its ultimate envisionment, VR causes the computer to become effectively invisible by taking over the human sensory and affector systems [Rheingold 91]. VR is extremely useful in scientific visualization and entertainment, and will be very significant for those niches. But as a tool for productively changing everyone’s relationship to computation, it has two crucial flaws: first, at the present time (1992), and probably for decades, it cannot produce a simulation of significant verisimilitude at reasonable cost (today, at any cost). This means that users will not be fooled and the computer will not be out of the way. Second, and most importantly, it has the goal of fooling the user — of leaving the everyday physical world behind. This is at odds with the goal of better integrating the computer into human activities, since humans are of and in the everyday world.

Ubiquitous computing is exploring quite different ground from Personal Digital Assistants, or the idea that computers should be autonomous agents that take on our goals. The difference can be characterized as follows. Suppose you want to lift a heavy object. You can call in your strong assistant to lift it for you, or you can be yourself made effortlessly, unconsciously, stronger and just lift it. There are times when both are good. Much of the past and current effort for better computers has been aimed at the former; ubiquitous computing aims at the latter.

The approach I took was to attempt the definition and construction of new computing artifacts for use in everyday life. I took my inspiration from the everyday objects found in offices and homes, in particular those objects whose purpose is to capture or convey information. The most ubiquitous current informational technology embodied in artifacts is the use of written symbols, primarily words, but including also pictographs, clocks, and other sorts of symbolic communication. Rather than attempting to reproduce these objects inside the virtual computer world, leading to another “desktop model” [Buxton 90], instead I wanted to put the new kind of computer also out in this world of concrete information conveyers. And because these written artifacts occur in many different sizes and shapes, with many different affordances, so I wanted the computer embodiments to be of many sizes and shapes, including tiny inexpensive ones that could bring computing to everyone.

The physical affordances in the world come in all sizes and shapes; for practical reasons our ubiquitous computing work begins with just three different sizes of devices: enough to give some scope, not enough to deter progress. The first size is the wall-sized interactive surface, analogous to the office whiteboard or the home magnet-covered refrigerator or bulletin board. The second size is the notepad, envisioned not as a personal computer but as analogous to scrap paper to be grabbed and used easily, with many in use by a person at once. The cluttered office desk or messy front hall table are real-life examples. Finally, the third size is the tiny computer, analogous to tiny individual notes or PostIts, and also like the tiny little displays of words found on book spines, lightswitches, and hallways. Again, I saw this not as a personal computer, but as a pervasive part of everyday life, with many active at all times. I called these three sizes of computers, respectively, boards, pads, and tabs, and adopted the slogan that, for each person in an office, there should be hundreds of tabs, tens of pads, and one or two boards. Specifications for some prototypes of these three sizes in use at PARC are shown in figure 1.

This then is phase I of ubiquitous computing: to construct, deploy, and learn from a computing environment consisting of tabs, pads, and boards. This is only phase I, because it is unlikely to achieve optimal invisibility. (Later phases are yet to be determined). But it is a start down the radical direction, for computer science, away from attention on the machine and back on the person and his or her life in the world of work, play, and home.

Hardware Prototypes

New hardware systems design for ubiquitous computing has been oriented towards experimental platforms for systems and applications of invisibility. New chips have been less important than combinations of existing components that create experimental opportunities. The first ubiquitous computing technology to be deployed was the Liveboard [Elrod 92], which is now a Xerox product. Two other important pieces of prototype hardware supporting our research at PARC are the Tab and the Pad.


The ParcTab is a tiny information doorway. For user interaction it has a pressure sensitive screen on top of the display, three buttons underneath the natural finger positions, and the ability to sense its position within a building. The display and touchpad it uses are standard commercial units.

The key hardware design problems in the pad are size and power consumption. With several dozens of these devices sitting around the office, in briefcases, in pockets, one cannot change their batteries every week. The PARC design uses the 8051 to control detailed interactions, and includes software that keeps power usage down. The major outboard components are a small analog/digital converter for the pressure sensitive screen, and analog sense circuitry for the IR receiver. Interestingly, although we have been approached by several chip manufacturers about our possible need for custom chips for the Tab, the Tab is not short of places to put chips. The display size leaves plenty of room, and the display thickness dominates total size. Off-the-shelf components are more than adequate for exploring this design space, even with our severe size, weight, and power constraints.

A key part of our design philosophy is to put devices in everyday use, not just demonstrate them. We can only use techniques suitable for quantity 100 replication, which excludes certain things that could make a huge difference, such as the integration of components onto the display surface itself. This technology, being explored at PARC, ISI, and TI, while very promising, is not yet ready for replication.

The Tab architecture is carefully balanced among display size, bandwidth, processing, and memory. For instance, the small display means that even the tiny processor is capable of four frame/sec video to it, and the IR bandwidth is capable of delivering this. The bandwidth is also such that the processor can actually time the pulse widths in software timing loops. Our current design has insufficient storage, and we are increasing the amount of non-volatile RAM in future tabs from 8k to 128k. The tab’s goal of postit-note-like casual use puts it into a design space generally unexplored in the commercial or research sector.


The pad is really a family of notebook-sized devices. Our initial pad, the ScratchPad, plugged into a Sun SBus card and provided an X-window-system-compatible writing and display surface. This same design was used inside our first wall-sized displays, the liveboards, as well. Our later untethered pad devices, the XPad and MPad, continued the system design principles of X-compatibility, ease of construction, and flexibility in software and hardware expansion.

As I write, at the end of 1992, commercial portable pen devices have been on the market for two years, although most of the early companies have now gone out of business. Why should a pioneering research lab be building its own such device? Each year we ask ourselves the same question, and so far three things always drive us to continue to design our own pad hardware.

First, we need the right balance of features; this is the essence of systems design. The commercial devices all aim at particular niches, and so balance their design to that niche. For research we need a rather different balance, all the more so for ubiquitous computing. For instance, can the device communicate simultaneously along multiple channels? Does the O.S support multiprocessing? What about the potential for high-speed tethering? Is there a high-quality pen? Is there a high-speed expansion port sufficient for video in and out? Is sound in/out and ISDN available? Optional keyboard? Any one commercial device tends to satisfy some of these, ignore others, and choose a balance of the ones it does satisfy that optimize its niche, rather than ubiquitous computing-style scrap computing. The balance for us emphasizes communication, ram, multi-media, and expansion ports.

Second, apart from balance are the requirements for particular features. Key among these are a pen emphasis, connection to research environments like Unix, and communication emphasis. A high-speed (>64kbps) wireless capability is built into no commercial devices, nor do they generally have a sufficiently high speed port to which such a radio can be added. Commercial devices generally come with DOS or Penpoint, and while we have developed in both, they are not our favorite research vehicles because of lack of full access and customizability.

The third thing driving our own pad designs is ease of expansion and modification. We need full hardware specs, complete O.S. source code, and the ability to rip-out and replace both hardware and software components. Naturally these goals are opposed to best price in a niche market, which orients the documentation to the end user, and which keeps price down by integrated rather than modular design.

We have now gone through three generations of Pad designs. Six scratchpads were built, three XPads, and thirteen MPads, the latest. The MPad uses an FPGA for almost all random logic, giving extreme flexibility. For instance, changing the power control functions, and adding high-quality sound, were relatively simple FPGA changes. The Mpad has built-in both IR (tab compatible) and radio communication, and includes sufficient uncommitted space for adding new circuit boards later. It can be used with a tether that provides it with recharging and operating power and an ethernet connection. The operating system is a standalone version of the public-domain Portable Common Runtime developed at PARC [Weiser 89].


FIGURE 1 – some hardware prototypes in use at PARC


FIGURE 2 – Photographs of each of tabs, pads, boards (at end of paper).


The CS of Ubicomp

In order to construct and deploy tabs, pads, and boards at PARC, we found ourselves needing to readdress some of the well-worked areas of existing computer science. The fruitfulness of ubiquitous computing for new Computer Science problems clinched our belief in the ubiquitous computing framework.

In what follows I walk up the levels of organization of a computer system, from hardware to application. For each level I describe one or two examples of computer science work required by ubiquitous computing. Ubicomp is not yet a coherent body of work, but consists of a few scattered communities. The point of this paper is to help others understand some of the new research challenges in ubiquitous computing, and inspire them to work on them. This is more akin to a tutorial than a survey, and necessarily selective.

The areas I discuss below are: hardware components (e.g. chips), network protocols, interaction substrates (e.g. software for screens and pens), applications, privacy, and computational methods.

Issues of hardware components

In addition to the new systems of tabs, pads, and boards, ubiquitous computing needs some new kinds of devices. Examples of three new kinds of hardware devices are: very low power computing, low-power high-bits/cubic-meter communication, and pen devices.

Low Power

In general the need for high performance has dominated the need for low power consumption in processor design. However, recognizing the new requirements of ubiquitous computing, a number of people have begun work in using additional chip area to reduce power rather than to increase performance [Lyon 93]. One key approach is to reduce the clocking frequency of their chips by increasing pipelining or parallelism. Then, by running the chips at reduced voltage, the effect is a net reduction in power, because power falls off as the square of the voltage while only about twice the area is needed to run at half the clock speed.


Power = CL * Vdd2 * f

  • where CL is the gate capacitance, Vdd the supply voltage, and f the clocking frequency.


This method of reducing power leads to two new areas of chip design: circuits that will run at low power, and architectures that sacrifice area for power over performance. The second requires some additional comment, because one might suppose that one would simply design the fastest possible chip, and then run it at reduced clock and voltage. However, as Lyon illustrates, circuits in chips designed for high speed generally fail to work at low voltages. Furthermore, attention to special circuits may permit operation over a much wider range of voltage operation, or achieve power savings via other special techniques, such as adiabatic switching [Lyon 93].


A wireless network capable of accommodating hundreds of high speed devices for every person is well beyond the commercial wireless systems planned even ten years out [Rush 92], which are aimed at one low speed (64kbps or voice) device per person. Most wireless work uses a figure of merit of bits/sec x range, and seeks to increase this product. We believe that a better figure of merit is bits/sec/meter3. This figure of merit causes the optimization of total bandwidth throughout a three-dimensional space, leading to design points of very tiny cellular systems.

Because we felt the commercial world was ignoring the proper figure of merit, we initiated our own small radio program. In 1989 we built spread-spectrum transceivers at 900Mhz, but found them difficult to build and adjust, and prone to noise and multipath interference. In 1990 we built direct frequency-shift-keyed transceivers also at 900Mhz, using very low power to be license-free. While much simpler, these transceivers had unexpectedly and unpredictably long range, causing mutual interference and multipath problems. In 1991 we designed and built our current radios, which use the near-field of the electromagnetic spectrum. The near-field has an effective fall-off of r6 in power, instead of the more usual r2, where r is the distance from the transmitter. At the proper levels this band does not require an FCC license, permits reuse of the same frequency over and over again in a building, has virtually no multipath or blocking effects, and permits transceivers that use extremely low power and low parts count. We have deployed a number of near-field radios within PARC.


A third new hardware component is the pen for very large displays. We needed pens that would work over a large area (at least 60″x40″), not require a tether, and work with back projection. These requirements are generated from the particular needs of large displays in ubiquitous computing — casual use, no training, naturalness, multiple people at once. No existing pens or touchpads could come close to these requirements. Therefore members of the Electronics and Imaging lab at PARC devised a new infrared pen. A camera-like device behind the screen senses the pen position, and information about the pen state (e.g. buttons) is modulated along the IR beam. The pens need not touch the screen, but can operate from several feet away. Considerable DSP and analog design work underlies making these pens effective components of the ubiquitous computing system [Elrod 92].

Network Protocols

Ubicomp changes the emphasis in networking in at least four areas: wireless media access, wide-bandwidth range, real-time capabilities for multimedia over standard networks, and packet routing.

A “media access” protocol provides access to a physical medium. Common media access methods in wired domains are collision detection and token-passing. These do not work unchanged in a wireless domain because not every device is assured of being able to hear every other device (this is called the “hidden terminal” problem). Furthermore, earlier wireless work used assumptions of complete autonomy, or a statically configured network, while ubiquitous computing requires a cellular topology, with mobile devices frequently coming on and off line. We have adapted a media access protocol called MACA, first described by Phil Karn [Karn 90], with some of our own modifications for fairness and efficiency.

The key idea of MACA is for the two stations desiring to communicate to first do a short handshake of Request-To-Send-N-bytes followed by Clear-To-Send-N-bytes. This exchange allows all other stations to hear that there is going to be traffic, and for how long they should remain quiet. Collisions, which are detected by timeouts, occur only during the short RTS packet.

Adapting MACA for ubiquitous computing use required considerable attention to fairness and real-time requirements. MACA (like the original ethernet) requires stations whose packets collide to backoff a random time and try again. If all stations but one backoff, that one can dominate the bandwidth. By requiring all stations to adapt the backoff parameter of their neighbors we create a much fairer allocation of bandwidth.

Some applications need guaranteed bandwidth for voice or video. We added a new packet type, NCTS(n) (Not Clear To Send), to suppress all other transmissions for (n) bytes. This packet is sufficient for a basestation to do effective bandwidth allocation among its mobile units. The solution is robust, in the sense that if the basestation stops allocating bandwidth then the system reverts to normal contention.

When a number of mobile units share a single basestation, that basestation may be a bottleneck for communication. For fairness, a basestation with N > 1 nonempty output queues needs to contend for bandwidth as though it were N stations. We therefore make the basestation contend just enough more aggressively that it is N times more likely to win a contention for media access.

Two other areas of networking research at PARC with ubiquitous computing implications are gigabit networks and real-time protocols. Gigabit-per-second speeds are important because of the increasing number of medium speed devices anticipated by ubiquitous computing, and the growing importance of real-time (multimedia) data. One hundred 256kbps portables per office implies a gigabit per group of forty offices, with all of PARC needing an aggregate of some five gigabits/sec. This has led us to do research into local-area ATM switches, in association with other gigabit networking projects [Lyles 92].

Real-time protocols are a new area of focus in packet-switched networks. Although real-time delivery has always been important in telephony, a few hundred milliseconds never mattered in typical packet-switched applications like telnet and file transfer. With the ubiquitous use of packet-switching, even for telephony using ATM, the need for real-time capable protocols has become urgent if the packet networks are going to support multi-media applications. Again in association with other members of the research community, PARC is exploring new protocols for enabling multimedia on the packet-switched internet [Clark 92].

The internet routing protocol, IP, has been in use for over ten years. However, neither it nor its OSI equivalent, CLNP, provides sufficient infrastructure for highly mobile devices. Both interpret fields in the network names of devices in order to route packets to the device. For instance, the “13” in IP name is interpreted to mean net 13, and network routers anywhere in the world are expected to know how to get a packet to net 13, and all devices whose name starts with 13 are expected to be on that network. This assumption fails as soon as a user of a net 13 mobile device takes her device on a visit to net 36 (Stanford). Changing the device name dynamically depending on location is no solution: higher level protocols like TCP assume that underlying names won’t change during the life of a connection, and a name change must be accompanied by informing the entire network of the change so that existing services can find the device.

A number of solutions have been proposed to this problem, among them Virtual IP from Sony [Teraoka 91], and Mobile IP from Columbia University [Ioannidis 93]. These solutions permit existing IP networks to interoperate transparently with roaming hosts. The key idea of all approaches is to add a second layer of IP address, the “real” address indicating location, to the existing fixed device address. Special routing nodes that forward packets to the right real address, and keep track of where this address is, are required for all approaches. The internet community has a working group looking at standards for this area (contact [email protected] for more information).

Interaction Substrates

Ubicomp has led us into looking at new substrates for interaction. I mention four here that span the space from virtual keyboards to protocols for window systems.

Pads have a tiny interaction area — too small for a keyboard, too small even for standard handprinting recognition. Handprinting has the further problem that it requires looking at what is written. Improvements in voice recognition are no panacea, because when other people are present voice will often be inappropriate. As one possible solution, we developed a method of touch-printing that uses only a tiny area and does not require looking. As drawbacks, our method requires a new printing alphabet to be memorized, and reaches only half the speed of a fast typist [Goldberg 93].

Liveboards have a huge interaction area, 400 times that of the tab. Using conventional pulldown or popup menus might require walking across the room to the appropriate button, a serious problem. We have developed methods of location-independent interaction by which even complex interactions can be popped up at any location. [Kurtenbach 93].

The X window system, although designed for network use, makes it difficult for windows to move once instantiated at a given X server. This is because the server retains considerable state about individual windows, and does not provide convenient ways to move that state. For instance, context and window IDs are determined solely by the server, and cannot be transferred to a new server, so that applications that depend upon knowing their value (almost all) will break if a window changes servers. However, in the ubiquitous computing world a user may be moving frequently from device to device, and wanting to bring windows along.

Christian Jacobi at PARC has implemented a new X toolkit that facilitates window migration. Applications need not be aware that they have moved from one screen to another; or if they like, they can be so informed with an upcall. We have written a number of applications on top of this toolkit, all of which can be “whistled up” over the network to follow the user from screen to screen. The author, for instance, frequently keeps a single program development and editing environment open for days at a time, migrating its windows back and forth from home to work and back each day.

A final window system problem is bandwidth. The bandwidth available to devices in ubiquitous computing can vary from kilobits/sec to gigabits/sec, and with window migration a single application may have to dynamically adjust to bandwidth over time. The X window system protocol was primarily developed for ethernet speeds, and most of the applications written in it were similarly tested at 10Mbps. To solve the problem of efficient X window use at lower bandwidth, the X consortium is sponsoring a “Low Bandwidth X” (LBX) working group to investigate new methods of lowering bandwidth. [Fulton 93].


Applications are of course the whole point of ubiquitous computing. Two examples of applications are locating people and shared drawing.

Ubicomp permits the location of people and objects in an environment. This was first pioneered by work at Olivetti Research Labs in Cambridge, England, in their Active Badge system [Want 92]. In ubiquitous computing we continue to extend this work, using it for video annotation, and updating dynamic maps. For instance, the picture below (figure 3) shows a portion of CSL early one morning, and the individual faces are the locations of people. This map is updated every few seconds, permitting quick locating of people, as well as quickly noticing a meeting one might want to go to (or where one can find a fresh pot of coffee).


Figure 3. Display of CSL activity from personal locators.


PARC, EuroPARC, and the Olivetti Research Center have built several different kinds of location servers. Generally these have two parts: a central database of information about location that can be quickly queried and dumped, and a group of servers that collect information about location and update the database. Information about location can be deduced from logins, or collected directly from an active badge system. The location database may be organized to dynamically notify clients, or simply to facilitate frequent polling.

Some example uses of location information are: automatic phone forwarding, locating an individual for a meeting, and watching general activity in a building to feel in touch with its cycles of activity (important for telecommuting).

PARC has investigated a number of shared meeting tools over the past decade, starting with the CoLab work [Stefik 87], and continuing with videodraw and commune [Tang 91]. Two new tools were developed for investigating problems in ubiquitous computing. The first is Tivoli [Pedersen 93], the second Slate, each based upon different implementation paradigms. First their similarities: they both emphasize pen-based drawing on a surface, they both accept scanned input and can print the results, they both can have several users at once operating independently on different or the same pages, they both support multiple pages. Tivoli has a sophisticated notion of a stroke as spline, and has a number of features making use of processing the contents and relationships among strokes. Tivoli also uses gestures as input control to select, move, and change the properties of objects on the screen. When multiple people use Tivoli each must be running a separate copy, and connect to the others. On the other hand, Slate is completely pixel based, simply drawing ink on the screen. Slate manages all the shared windows for all participants, as long as they are running an X window server, so its aggregate resource use can be much lower than Tivoli, and it is easier to setup with large numbers of participants. In practice we have used slate from a Sun to support shared drawing with users on Macs and PCs. Both Slate and Tivoli have received regular use at PARC.

Shared drawing tools are a topic at many places. For instance, Bellcore has a toolkit for building shared tools [Hill 93], and Jacobsen at LBL uses multicast packets to reduce bandwidth during shared tool use. There are some commercial products [Chatterjee 92], but these are usually not multi-page and so not really suitable for creating documents or interacting over the course of a whole meeting. The optimal shared drawing tool has not been built. For its user interface, there remain issues such as multiple cursors or one, gestures or not, and using an ink or a character recognition model of pen input. For its substrate, is it better to have a single application with multiple windows, or many applications independently connected? Is packet-multicast a good substrate to use? What would it take to support shared drawing among 50 people, 5,000 people? The answers are likely both technological and social.

Three new kinds of applications of ubiquitous computing are beginning to be explored at PARC. One is to take advantage of true invisibility, literally hiding machines in the walls. An example is the Responsive Environment project led by Scott Elrod. This aims to make a building’s heat, light, and power more responsive to individually customized needs, saving energy and making a more comfortable environment.

A second new approach is to use so-called “virtual communities” via the technology of MUDs. A MUD, or “Multi-User Dungeon,” is a program that accepts network connections from multiple simultaneous users and provides access to a shared database of “rooms”, “exits”, and other objects. MUDs have existed for about ten years, being used almost exclusively for recreational purposes. However, the simple technology of MUDs should also be useful in other, non-recreational applications, providing a casual environment integrating virtual and real worlds [Curtis 92].

A third new approach is the use of collaboration to specify information filtering. Described in the December 1992 issue of Communcations of the ACM, this work by Doug Terry extends previous notions of information filters by permitting filters to reference other filters, or to depend upon the values of multiple messages. For instance, one can select all messages that have been replied to by Smith (these messages do not even mention Smith, of course), or all messages that three other people found interesting. Implementing this required inventing the idea of a “continuous query”, which can effectively sample a changing database at all points in time. Called “Tapestry”, this system provides new ways for people to invisibly collaborate.

Privacy of Location

Cellular systems inherently need to know the location of devices and their use in order to properly route information. For instance, the traveling pattern of a frequent cellular phone user can be deduced from the roaming data of cellular service providers. This problem could be much worse in ubiquitous computing with its more extensive use of cellular wireless. So a key problem with ubiquitous computing is preserving privacy of location. One solution, a central database of location information, means that the privacy controls can be centralized and so perhaps done well — on the other hand one break-in there reveals all, and centrality is unlikely to scale worldwide. A second source of insecurity is the transmission of the location information to a central site. This site is the obvious place to try to snoop packets, or even to use traffic analysis on source addresses.

Our initial designs were all central, initially with unrestricted access, gradually moving towards controls by individual users on who can access information about them. Our preferred design avoids a central repository, but instead stores information about each person at that person’s PC or workstation. Programs that need to know a person’s location must query the PC, and run whatever gauntlet of security the user has chosen to install there. EuroPARC uses a system of this sort.

Accumulating information about individuals over long periods is both one of the more useful things to do, and also most quickly raises hackles. A key problem for location is how to provide occasional location information for clients that need it while somehow preventing the reliable accumulation of long-term trends about an individual. So far at PARC we have experimented only with short-term accumulation of information to produce automatic daily diaries of activity [Newman 90].

It is important to realize that there can never be a purely technological solution to privacy, that social issues must be considered in their own right. In the computer science lab we are trying to construct systems that are privacy enabled, that can give power to the individual. But only society can cause the right system to be used. To help prevent future oppressive employers or governments from taking this power away, we are also encouraging the wide dissimenation of information about location systems and their potential for harm. We have cooperated with a number of articles in the San Jose Mercury News, the Washington Post, and the New York Times on this topic. The result, we hope, is technological enablement combined with an informed populace that cannot be tricked in the name of technology.

Computational Methods

An example of a new problem in theoretical computer science emerging from ubiquitous computing is optimal cache sharing. This problem originally arose in discussions of optimal disk cache design for portable computer architectures. Bandwidth to the portable machine may be quite low, while its processing power is relatively high, introducing as a possible design point the compression of pages in a ram cache, rather than writing them all the way back over a slow link. The question arises of the optimal strategy for partitioning memory between compressed and uncompressed pages.

This problem can be generalized as follows [Bern 93]:

The Cache Sharing Problem. A problem instance is given by a sequence of page requests. Pages are of two types, U and C (for uncompressed and compressed), and each page is either IN or OUT. A request is served by changing the requested page to IN if it is currently OUT. Initially all pages are OUT. The cost to change a type-U (type-C) page from OUT to IN is CU (respectively, CC). When a requested page is OUT, we say that the algorithm missed. Removing a page from memory is free.

Lower Bound Theorem: No deterministic, on-line algorithm for cache sharing can be c-competitive for

  • c < MAX (1+CU/(CU+CC), 1+CC/(CU+CC))

This lower bound for c ranges from 1.5 to 2, and no on-line algorithm can approach closer to the optimum than this factor. Bern et al also construct an algorithm that achieves this factor, therefore providing an upper bound as well. They further propose a set of more general symbolic programming tools for solving competitive algorithms of this sort.

Concluding remarks

As we start to put tabs, pads, and boards into use, phase I of ubiquitous computing should enter its most productive period. With this substrate in place we can make much more progress both in evaluating our technologies and in choosing our next steps. A key part of this evaluation is using the analyses of psychologists, anthropologists, application writers, artists, marketers, and customers. We believe they will find some things right; we know they will find some things wrong. Thus we will begin again the cycle of cross-disciplinary fertilization and learning. Ubicomp seems likely to provide a framework for interesting and productive work for many more years or decades, but we have much to learn about the details.

Acknowledgements: This work was funded by Xerox PARC. Portions of this work were sponsored by DARPA. Ubiquitous computing is only a small part of the work going on at PARC; we are grateful for PARC’s rich, cooperative, and fertile environment in support of the document company.Bern 93. Bern, M., Greene, D., Raghunathan. On-line algorithms for cache sharing. 25th ACM Symposium on Theory of Computing, San Diego, 1993.

Buxton 90. Buxton, W. (1990). Smoke and Mirrors. Byte,15(7), July 1990. 205-210.

Chatterjee 92. Chatterjee, Shalini. Sun enters computer conferencing market. Sunworld. pp. 32-34. Vol 5, no. 10. October 1992. Integrated Media, San Francisco.

Clark 92. Clark, David D., Shenker, Scott, Zhang, Lixia. Supporting real-time applications in an integrated services packet network:architecture and mechanism. SIGCOMM ’92 Conference Proceedings. Communicatins architectures and protcosl. August 17-20, 1992. Baltimore, Maryland. Computer Communication Review. Vol. 22, no. 4, October 1992.published by Accosication for Comptuing Machinery, New York, NY. pp. 14-26

Curtis 92. Curtis, Pavel. MUDDING: social phenomena in text-based virtual realities. DIAC – Directions and Implications of Advanced Computing. May, 1992 Symposium Proceedings. Computer Professionals for Social Responsibility. Palo Alto, CA.

Elrod 92. Elrod, Bruce, Gold, Goldberg, Halasz, Janssen, Lee, McCall, Pedersen, Pier, Tang, and Welch. Liveboard: a large interactive display supporting group meetings, presentations and remote collaboration. pp. 599-607. CHI ’92 Conference proceedings. May 1992. ACM, New York, NY.

Fulton 93. Fulton, Jim and Kantarjiev, Chris. An Update on Low Bandwidth X (LBX). Proceedings of the Seventh Annual X Technical Conference, January, 1993, Boston, MA. Published in The X Resource by O’Reilly and Associates. (to appear)

Goldberg 93. Goldberg, David, Richardson, Cate. Touch Typing with a Stylus. to appear, INTERCHI ’93.

Hill 93. Hill, R.D. , T. Brinck, J.F. Patterson, S.L. Rohall, and W.T. Wilner. The RENDEZVOUS Language and Architecture: Tools for Constructing Multi-User Interactive Systems. Communications of the ACM, Vol. 36, No. 1 (Jan. 1993). (to appear)

Ioannidis 93. Ioannidis, John, Maguire, Gerald Q., Jr. The Design and Implementation of a Mobile Internetworking Architecture. Usenix Conference Proceedings, Usenix ’93. January 1993. to appear.

Karn 90. Karn, P. MACA – A New Channel Access Method for Packet Radio. Proceedings of the ARRL 9th Computer Networking Conference, London Ontario, Canada, September 22, 1990. ISBN 0-87259-337-1.

Kay 91. Kay, Alan. Computers, Networks, and Education. Scientific American, September 1991. pp. 138-148.

Kurtenbach 93. Kurtenbach, Gordon, Buxton, William. The Limits of Expert Performance Using Hierarchic Marking Menus. to appear, INTERCHI ’93

Lave 91. Lave, Jean. Situated learning: legitimate peripheral participation. Cambridge University Press. Cambridge. New York, NY. 1991.

Lyles 92. Lyles, B. J., Swinehart, D. C. The emerging gigabit environment and the role of local ATM. IEEE Communications Magazine. April 1992. p. 52-57.

Lyon 93. Lyon, Richard F. Cost, Power, and Parallelism in Speech Signal Processing. Custom Integrated Circuits Conference, IEEE, 1993. May 9-12, 1993. San Diego.

Newman 90. Newman, William M., Eldridge, Margery A., Lamming, Machael G. PEPYS: Generating Autobiographies by Automatic Tracking. EuroPARC technical report. Cambridge England.

Pedersen 93. Pedersen, Elin, McCall, Kim, Moran, Thomas P., Halasz, Frank G. Tivoli: An Electronic Whiteboard for Informal Workgroup Meetings. to appear, INTERCHI ’93

Rheingold 91. Rheingold, Howard. Virtual Reality. Summit Books. New York, NY. 1991.

Rush 92. Rush, Charles M. How WARC ’92 Will Affect Mobile Services. IEEE Communications Magazine. October 1992. pp. 90-96.

Stefik 87. Stefik, M, Foster, G., Bobrow, D.G., Kahn, K., Lanning, S., and Suchman, L. Beyond the chalkboard: computer support for collaboration and problem solving in meetings. CACM 30, 1. January 1987. pp. 32-47.

Suchman 85. Suchman, Lucy A. Plans and Situated Actions: The problem of human-machine communication. Xerox PARC Technical Report ISL-6. February 1985

Tang 91. Tang, John C., Minneman, Scott L. VideoDraw: A video interface for collaborative drawing. ACM Trans. on Office Information Systems. Vol 9, no 2. April 1991. pp. 170-184.

Teraoka 91. Teraoka, Fumio, Tokote, Yasuhiko, Tokoro, Mario. A network architecture providing host migration transparency. Proceedings of SIGCOMM’91. ACM pp. 209-220. September 1991.

Tesler 91. Tesler, Lawrence G. Networked Computing in the 1990’s. Scientific American, September 1991. pp. 86-93.

Want 92a. Want, Roy, Hopper, Andy, Falcao, Veronica, and Gibbons, Jonathan. The active badge location system. ACM T. on Information Systems. Vol 10, No. 1 January 1992. pp. 91-102.

Weiser 89. Weiser, Mark, Demers, Alan, Hauser, C. The Portable Common Runtime Approach to Interoperability. Proceedings of the ACM Symposium on Operating Systems Principles, December 1989.

Weiser 91. Weiser, Mark. The Computer for the Twenty-First Century. Scientific American. September 1991. pp. 94-104.

This is an archive of Mark Weiser’s ubiquitous computing website ( which disappeared from the internet in 2018 some time after Mark Weiser passed away. We wanted to preserve Mark Weiser’s knowledge about ubiquitous computing and are permanently hosting a selection of important pages from

Access Time

One of the most commonly quoted performance statistics for CD-ROM drives is access time. As with most commonly-used performance metric, it is abused at least as much as it is used properly. Curiously, access time is used extensively in quoting the specs of CD-ROM drives, but is virtually never mentioned with respect to hard disks. With hard drives it is much more common to see quotes of the other metrics that are combined to make up access time. (This despite how similarly the devices access data…).

Access time is meant to represent the amount of time it takes from the start of a random read operation until the data starts to be read from the disk. It is a composite metric, really being composed of the following other metrics:

  • Speed Change Time: For CLV drives, the time for the spindle motor to change to the correct speed.
  • Seek Time: The time for the drive to move the heads to the right location on the disk.
  • Latency: The amount of time for the disk to turn so that the right information spins under the read head.

Although access time is made up of the time for these separate operations, this doesn’t mean that you can simply add these other measurements together to get access time. The relationship is more complex than this because some of these items can happen in parallel. For example, there is no reason that the speed of the spindle motor couldn’t be varied at the same time that the heads are moved (and in fact this is done).

The access time of CD-ROM drives in general depends on the rated “X” speed of the drive, although this can and does vary widely from drive to drive. The oldest 1X drives generally had truly abysmal access times, often exceeding 300 ms; as drives have become faster and faster, access times have dropped, and now are below 100 ms on the top-end drives.

Note that while faster “X” rated drives have lower access times, this is due to improvements that reduce the three metrics listed above that contribute to access time. Some of it (latency for example) is reduced when you spin the disk at 8X instead of 1X. On the other hand, seek time improvement is independent of the spin speed of the disk, which is why some 8X drives will have much better access time performance than other 8X drives, for example.

Even the fastest CD-ROM drives are significantly slower than even the slowest hard disks; access time on a high-end CD-ROM is still going to be four or five times higher than that of a high-end disk drive. This is just the nature of the device; CD-ROM drives are based on technology originally developed for playing audio CDs, where random seek performance is very unimportant. CDs do not have cylinders like a hard disk platter, but rather a long continuous spiral of bits, which makes finding specific pieces of data much more difficult.

Even though access time is important in some ways, its importance is generally vastly overstated by the people that sell CD-ROM drives. Random access performance is one component of overall CD-ROM performance, and how essential it is depends on what you are doing with your drive. However even if high random access performance is important, you must bear in mind that there are far fewer random reads done, in general, to a CD than to a hard disk.

Another point is that manufacturers are not always consistent in how they define their averages. Some companies may use different testing methods, and some may even exaggerate in order to make their drives look much better than they actually are. A small difference in quoted access time is not usually going to make any noticeable real-world difference. For most purposes, a drive with a 100 ms access time is going to behave the same as one with a 110 ms access time. It’s usually better at that point to differentiate them based on other performance characteristics or features (or price).

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.

101-Key “Enhanced” Keyboard Layout

In 1986, IBM introduced the IBM PC/AT Model 339. Included in this last AT-family system was the new Enhanced 101-key keyboard. Little did IBM realize at the time, perhaps, but this 101-key keyboard would become the de-facto standard for keyboards for the next decade and beyond. Even today’s Windows keyboards and fancy variants with extra buttons and keys are based on this layout.

101-key "enhanced" keyboard

The “Enhanced” keyboard was electrically the same as the 84-key AT keyboard, but featured a radically redesigned key layout. The major changes included these:

  • Dedicated Cursor and Navigation Keys: Finally, separate keys were provided for cursor control and navigation. This enabled the numeric keyboard to be used along with the cursor and navigation keys. The cursor keys were also made into an “inverted-T” configuration for easier switching between “Up” and “Down” with a single finger.
  • Relocated Function Keys: The function keys were moved from the left-hand side of the keyboard to a row along the top, and divided into groups of four for convenience. While many users had been asking for this, they found that sometimes the grass really isn’t greener on the other side of the fence, as I discuss below…
  • Relocated <Esc> and <Caps Lock> Keys: The <Esc> key was moved back to the left-hand side of the keyboard, though it was placed up above the main typing area. The <Caps Lock> key was moved above the left <Shift> key.
  • Extra Function Keys: Two additional function keys, <F11> and <F12> were added to the keyboard.
  • Extra <Ctrl> and <Alt> Keys: Additional <Ctrl> and <Alt> keys were added on the right side of the <Space Bar>.
  • Extra Numeric Keypad Keys: The numeric keypad was fitted with an additional <Enter> key, as well as the “/” (divide operator) that had been missing up to that point.

Compared the 84-key keyboard the Enhanced keyboard layout was perceived by most users to be far superior. It was an immediate hit despite its one obvious inferiority to the AT keyboard: the smaller main <Enter> key. (The <Space Bar> is also a bit smaller.) Obviously, some of the changes made with the Enhanced keyboard are undeniable. However, others are in this author’s opinion good examples of the old warning: “be careful what you ask for”…

Many PC users, after having complained for years about changes they wanted made to the PC keyboard layout, found they weren’t all that happy with them once their wish was granted! Having never complained about the issues that were changed with the Enhanced keyboard myself, I found some of the changes quite frustrating–and I later discovered that I was not alone. My personal beefs with this layout involve the locations of the following:

  • Left <Ctrl> Key: With the older layout, the left-hand <Ctrl> key is readily accessible, and it is used by computer enthusiasts dozens, if not hundreds of times a day. (For example, cut, copy and paste are universal functions with standard Windows short-cuts of <Ctrl>+X, <Ctrl>+C and <Ctrl>+V respectively.) The new design puts the <Ctrl> key below the main keyboard, requiring a move of the entire left hand to reach it. And while having the <Caps Lock> key above the left <Shift> may be of use to some, I use the <Caps Lock> key maybe once or twice a month, how about you? :^) Overall, a really bad swap in my opinion.
  • Function Keys: Having the function keys on the left-side of the keyboard makes them easy to reach, particularly in combination with the <Shift>, <Ctrl> and <Alt> keys. Again, these are frequently used keys which are hard to reach when above the keyboard; most combinations that used to be simple with one hand now require two. For example, a command I use frequently when writing is <Ctrl>+<F6>, the Microsoft Word (and FrontPage) function to switch between documents. Compare the motion required to type this combination on an Enhanced keyboard to what was required with the function keys on the left side and the <Ctrl> key above the <Shift> key. Also consider <Alt>+<F4>, the standard combination to close a Windows application… and so it goes.
    The real irony, of course, is that the “on-screen labels corresponding to function keys”, which is what caused people to want the function keys along the top of the keyboard, disappeared from software applications many years ago!
  • <Esc> Key: This key is still a reach with the Enhanced design. Compare how often you use the <Esc> key in a day to the number of times you type a backwards quote or tilde! Again, a poorly-considered decision.

Despite these limitations, the 101-key keyboard remains the standard (actually, the 104-key Windows keyboard is the standard now, but the two layouts are nearly identical). Of course, countless variations of the basic design exist. A common modification is to enlarge the <Enter> key back to its “84-key layout size”, and squeeze the backslash / vertical-pipe key between the “=/+” key and the <Backspace>. An improvement in my estimation!

As for me, rather than curse the darkness, I lit a candle: I use a 124-key Gateway Anykey programmable keyboard with function keys both above and to the left of the main typing area, and a large main <Enter> key. I relocate the left <Ctrl> to where it belongs and the <Caps Lock> key somewhere out of the way where it belongs. :^) I swap the <Esc> key and the backquote/tilde key as well. Ah, freedom. :^)

The PC Guide
Site Version: 2.2.0 – Version Date: April 17, 2001
© Copyright 1997-2004 Charles M. Kozierok. All Rights Reserved.

This is an archive of Charles M. Kozierok’s PCGuide ( which disappeared from the internet in 2018. We wanted to preserve Charles M. Kozierok’s knowledge about computers and are permanently hosting a selection of important pages from PCGuide.