The PARCTAB system has been in use since March 1993 and now serves a small community of users. We have made a number of useful observations during this period and have begun to understand its successes and failures.
PARC was a convenient test site for the PARCTAB system because installation was very easy. Before the project began every office already contained a workstation connected by an ethernet. The hallways and common areas also had access to nearby workstations. It was easy to install a communication cell in an office by using velcro to attach a transceiver to the ceiling and then to run phone cable down a wall into a junction box. The junction box usually rests on the floor under a desk and has a power cable, and connects to the RS232 port of the workstation. Typically, the installation takes about 15 minutes.
Some users also installed cells in their homes. They already had ISDN lines, which connect a home ethernet to the office network, and so a transceiver connected to a workstation at home was effectively tied to the PARCTAB infrastructure.
The first PARCTAB system released in March '93 consisted of 20 users and 25 cells. The experience gained in this time enabled a second release in April '94. The latter system was somewhat larger with a community of about 41 users and 50 cells. It included many improvements that enhanced the performance of the communication channel and the tabs' perceived reliability.
For example, the original system relied on a central name-and-maintain service (see Section 5.2) to route packets to tabs; when the service was unavailable the PARCTAB system could not function. The new release has a distributed name service that uses a network multicast mechanism to determine the address of system components.
We discovered in the first release there were problems caused by high utilization of the infrared network. High loads cause three problems: infrared packets are more likely to be corrupted; transmit buffers in the transceiver overflow, causing packets to be dropped; and the corrupted and dropped packets caused more retransmissions, increasing the load. The high load exposed bugs in the system design and implementation such as race conditions and badly-tuned retransmission policies.
To improve user's confidence in the system, we had to increase its reliability and availability. This involved not only fixing bugs but also mundane improvements such as a low-battery indicator for tabs. System components also needed mechanisms for self monitoring. All the PARCTAB system processes now have control panels designed to provide information in the event of a failure. We have also put new mechanisms in place to monitor and maintain the IR-gateway and the agent processes.
The PARCTAB could not be used effectively in several rooms in our building because of IR noise due to fluorescent lamps controlled by electronic ballasts. This is a waste of a unique form of communication bandwidth. Unfortunately, electronic ballasts are slowly replacing the older magnetic ballasts because they are more energy efficient. We found a considerable variation in interference levels from lamps made by different manufacturers. Some produce acceptable levels of IR, and it would be useful if lamp manufacturers were required to adhere to a maximum limit for IR emissions.
Positioning of a room transceiver is also important. Installers should avoid direct sunlight, that can change position throughout a day (and during the year), and proximity to fluorescent lamps and to obstructions on the ceiling. Transceivers in adjacent cells should be positioned carefully so that their signals do not pass through doorway or interior windows and cause interference.
Part of the benefit of building a real system has been the opportunity to study how a versatile personal information-terminal might be used in advance of a commercial system. We studied the 1994 release of the tab system for three months to determine its use characteristics. The participants all consented to automatic logging of system events.
We began recording two weeks after system deployment so that users
could familiarize themselves with the PARCTAB . To limit the data
to a manageable quantity, we logged only the following events:
Interactive, Switch, Idle, and Missing
.
Interactive occurs when a user powers up a tab, Switch
occurs when a user switches to a new application, Idle is generated
when a tab has not been used for 4 minutes, and Missing is a timeout
event generated by the system when the infrared network cannot detect
a particular tab. Each event was recorded along with a timestamp and
cell location. In addition, there were two questionnaires given out
to our users, one at the outset of the tab use study and one at the
close. This provided contextual information, and information to
interpret the logging data.
The switch events can be used to determine the relative popularity of the various PARCTAB applications. Figure 11 shows the percentage of invocations accounted for by each application. Four were distinctly more popular than the rest: e-mail, weather, file browser, and the loader. Possible implications of these results are discussed in Section 9.
Figure 11: Histogram showing the number of invocations for each application
(not including the shell or tshell) expressed as a percentage of the
total invocations of these applications during the test period.
Normalized results only count one invocation per day per user to
remove distortions that might arise when users experiment with an
application several times during a brief period. Applications that
might normally be invoked several times a day suffer under this
measure.
Another way of looking at application popularity is to consider how long each application was in use (see Figure 12). It should be noted that the total application interaction time is 4871 minutes over 3 months (13 weeks) for 41 users. This amounts to only 119 minutes/user or about 1.8 minutes/user/day (65 days, excluding weekends). From our logs the total number of application switches for all tabs throughout the study was 2996 and therefore the average interaction time was about 97 seconds.
The application popularity ranking is somewhat different from Figure 11. The e-mailer, unistroke test and learn program, unistroke notetaker, file browser, and the loader are the most long-lived applications. The weather program falls to 8th place (perhaps because it only imparts a small amount of information at any one time). Meanwhile the note-taker moves up to 3rd place -- not surprising, as taking notes is by its nature a time-consuming activity. It is interesting to observe that reading e-mail, browsing system files, and loading data turn out to be the most used in both measurements.
This use pattern differed from the participants own expectations of use. Although they expected to read e-mail, (four of the participants did not use e-mail on the tab at all, due to incompatible mail systems), over half commented that they expected to use the tab primarily as a calendar. It is also worth noting that according to user reports the e-mail program was used to read e-mail much more than to send e-mail using Unistrokes. The Unistroke test and learn programs appear in the ranking even though they are typically not activated very often; users may spend a block of time running them when first acquiring the skill.
Figure 12: Histogram showing the total interaction time by users for each
application in the tab system during the 3 month test period (not-including
the shell, 1273 minutes, and the tshell, 1081 minutes).
Graph 13 shows the percentage of application interactions that last less than a given time. We have removed interactions of less than 10 seconds because users often turn a tab on and then off immediately to confirm that it is working normally. From this graph we can see that 50% of interactions last less than 100 seconds (1.7 mins), 75% less than 230 seconds (3.8 mins) and 90% less than 500 seconds (8.3 mins). This supports our notion of the tab as a device for ``casual'' interactions.
Figure 13: Graph showing the percentage of application interactions that
were under a given time during the test period.
Figure 14 shows what fraction of users had their tabs turned on for various total periods of time. The study group can be roughly divided into three user types. 7% (3 people) used the tab for 360-480 minutes during the test (6.4 minutes/day). 15% (6 people) used it for 144-360 minutes (3.9 minutes/day) and 78% (32 people) used it for less than 144 minutes in total (1.1 minutes/day). The average use time for the majority was very small, implying their interactions were generally very brief.
Figure 14: Histogram showing the number of users against their total interaction
time divided into 20 equal divisions.
Figure 15 shows interaction time for each user, subdivided according to location: in their own office (black); in a common area such as a conference room, tea area or seminar room (grey); or in a hall or another person's office (white). Only 3 people used a tab primarily (for more than 50% of their total interaction time) in somebody else's office. Approximately 61% (25 people) of our community used the tab primarily in their own rooms, and 27% (11 people) used it primarily in a common area. Interestingly enough, for each pattern of use the preference was quite clear.
Figure 15: Histogram showing the total interaction time for each user in
seconds split between three location types: a user's own office, a
common area, a hall or another person's office.
By pooling the results of Figure 15 we can determine that people used tabs in their own offices 57% of the time, in a common area 32% of the time , and in another office 11% of the time (see Figure 16). 7% of own-office interactions are in the presence of other tabs. 90% of common area interactions and 85% of other-office interactions are also in this category.
The multiple-user applications, group drawing and remote pointing, were not available for the duration of the use study. Group applications like this would have generated a much higher network-load in the common areas, but are likely uses of a ubiquitous mobile device.
Figure 15 shows that there is not a typical use pattern among the study group. Our questionnaires showed that there were as many different expectations of the tab system as there were participants in the study. For example, researchers developing applications on the tab that expected to use the tab a great deal did not necessarily have the largest interactions times, even though they had to use the tab for their daily work. In contrast, some researchers who did not expect to use the tab found that visitor demonstrations of the device added significantly to their total usage time.
Figure 16: Histogram showing the total interaction time by all users for each
of the three general areas: a user's own office, a common area, a hall
or another person's office.
These results are important for overall system design because multiple tabs interacting in the same area have a strong impact on the available bandwidth. The PARCTAB system needs to be able to handle a usage pattern in which at least 42% of all interactions occur with multiple tabs present.
Although the previous graphs give an indication of the way the tab was used it is important to acknowledge the limitations of this study in representing the use of the tab as a consumer item. First, the the user group was too small for statistically significant results. Second, the system was still under development and the applications were not fully supported. Furthermore, participants in the study were not customers but rather laboratory staff using the tab as a prototype. It was up to them to invent ways to use the tab, develop new applications and create ways to incorporate the tab into established work patterns. As a result, we must qualify the numbers with anecdotal evidence and further discussion of the ways people used the tab. Some of these remarks are listed below:
Rich Gold: does not see any value in using a tab in his own office because a powerful workstation is at hand.
John Ellis: prefers to use the tab in his own office to read his e-mail so that he does not have to rearrange the windows on his workstation screen.
Dan Swinehart: found the tab system had a long response time, but found that the tab system was faster than Mosaic for finding the definition of a word, .
Helen Davis: has used the email application and Unistrokes to take notes during seminars and then mailed them to herself.
A number of people found the PARCTAB too heavy or awkward to wear.
Two women tab users (Karin Petersen and Nancy Freige) remarked that the design of the belt clip was oriented towards a particular clothing style. For example, not all outfits include belts, and furthermore not all belts work well with clip on devices. Doug Terry also found the tab clip inadequate for his use. Instead he used a small zippered nylon and (infrared transparent) fishnet pouch to hold a tab so that it could be attached to his belt and continue to report his location.
A researcher who preferred to remain anonymous commented on the difficulties of building new applications in Modula-3: `I don't want to say anything against Modula-3 but if I have to learn a new language at the same time as trying to program a new [computer] I may not get much done.'
The ease of reading text on the small screen surprised most of the participants in the use study. At the beginning of the study we found almost 1/2 of the participants had commented that because of the low resolution of the screen they did not intend to read longer files.
As the list above indicates, it is difficult to suggest a `typical' use of the PARCTAB The PARCTAB system was an experiment that many people volunteered to participate in. It was shaped by their own ideas, needs and contributions. A direct consequence of building a system that can be used by a community is that it is possible to gain understanding of the real problems (see Section 9), issues to be addressed, and activities that need to be supported.
To gain more general experience we gave the tab system (including tabs, transceivers, and software) to a number of other research departments. The largest of these sites was the Rank Xerox Research Centre (Cambridge, England) with 12 transceivers and 10 PARCTAB s. Flinders University (Adelaide, Australia) University of Washington, University of Toronto and Olivetti Research Ltd (Cambridge, England) also received small numbers of PARCTAB system components for their own research. RXRC produced a number of applications (see Section 7), and the University of Toronto now uses tabs to control the equipment in its ``telepresence'' room.