Appendix 3: Performance Graphs

The afternoon of 17 February 1995, PowerWeb was tested at the Marcus Whitman Elementary School Macintosh lab in Richland, WA. A plug-in was developed that performed a given number of instructions, while using all connected PowerWeb Slaves to the best of its ability. These graphs are an indication of how well PowerWeb performed in the real world; each PowerWeb Slave was assigned to use 50% of the CPU.

The first graph, Time, shows the total time (in sixtieths of a second) that it took to complete the job based on the number of connected PowerWeb Slaves.

Note how drastically the time drops off as additional PowerWeb Slaves are added to the system.

The next graph, Performance, is simply the inverse of time, calibrated so a system with one PowerWeb Slave corresponds to one.

It is quite linear based on the number of PowerWeb Slaves attached.

Finally, the efficiency can be computed with the assumption that efficiency tends to drop from 100% efficiency with zero slaves. This had to be extrapolated from our given information; here is the efficiency graph.

Note that the efficiency is not perfectly linear. Our hypothesis is that glitches occurred when faster and slower Macintoshes came online at different times; this would increase or decrease performance of the entire system.

Later, we reproduced some of our testing with four PowerWeb Slaves online; this time, however, we varied the percentage CPU allowed for PowerWeb use. Here is a graph with the new values superimposed on the old graph of time used.

Note that refers to the new points (at 10% and 90% of full CPU usage).

The performance graph shows, perhaps more clearly, the impact of adjusting the CPU usage parameter:

Again, note that refers to the new data points at 10% and 90% of full CPU usage.