cunningh@noscvax.UUCP (12/07/83)
This is a capsule summary of an article that caught my eye in the November '83 issue of CACM describing an interesting experiment on the results of delays in computer response time. "System response time[,] operator productivity, and job satisfaction" was written by Ray Barber and Henry Lucas (both at New York University when they wrote it). The experiment was conducted in the "circuit layout" organization of an unidentified large midwestern telephone company. The unknowing experimental subjects were over 100 clerks using five on-line systems (with a total transaction volume of 70,000+ per day) to develop and issue "circuit orders" (instructions to make specific connections disconnections or modifications to interstate telephone circuits. On an isolated group of 19 of the over 100 terminals in use, response time was deliberately degraded for four days. Neither the clerks nor their supervisors were told what was happening. The rest of the clerks formed the control group. In addition to collecting data on transaction volume, active time, response time and errors, attitudes questionaires were filled out by the clerks 12 days before the response-time slowdown, during the slowdowm, and 8 weeks afterwards. Over 1174 person-days worth of data were gathered during the slowdown. The objective of the study was, of course, to find out what happens when computer resonse time slows down. The results are interesting, even though the work done on the computer was definitely transaction processing, not text editing or similar work. The types of transactions these clerks did, however, is certainly more complex than doing airline reservations or simple bank-type transactions. Most, if not all of the transactions seem to involve queries and modifications on several (distributed) data bases. In order to complete the order, each clerk had to do about 8 significantly different operations, including one that generated a telegram to telephone engineers in a remote city. Often, they had to intersperse consultations with real people (phone company engineers) between their computer entry operations. It's no surprise, then, that the average (control) transaction time was 6 seconds (with considerable variability). The average slowed-down time was set at 14 seconds. The authors performed some fairly sophisticated statistical analyses, and their results seem pretty convincing. One result, intuitively obvious, is that the total transactions performed turned out as an inverse, roughly linear, function of response time. If the system allowed people to work faster, they did. Cursiously, small increases in delay time (in the 4 to 12 second range) actually seemed to produce less errors. Apparently operators used extra care so they wouldn't incure the longer delays twice. It was only with delays of over 12 seconds that the percentage of errors started to increase. Even more curiously, relatively long delays (20+ seconds), produced about as many errors as productive transactions. Seemingly because the clerks became confused, distracted or lost their place. The net effect is that for delays below six seconds or so, the number of productive transactions was roughly constant. Longer delays (with less errors) showed a gentle dropoff in productivity just because of the slowness. Beyond 12 seconds delay, productivity fell off drastically as errors increased. Naturally, job satisfaction generally declined as delay went up. However, the clerks' "interpersonal and supervisory relation" satisfaction increased. Apparently they enjoyed talking with their co-workers and supervisors more, even though a lot of the conversation seemed to be concerned with complaining about the bad computer performance. -- Bob Cunningham ..sdcsvax!noscvax!cunningh 21 17' 35" N 157 49' 38" W MILNET: cunningh@nosc-cc