This preview shows page 1. Sign up to view the full content.
Unformatted text preview: We show the results on two different machines to illustrate the effect of the time resolution on accuracy. The measurements on a Windows-NT system show characteristics similar to those we found for Linux using times (Figure 9.8). Since gettimeofday is implemented using the process timers, the error can be negative or positive, and it is especially erratic for short duration measurements. The accuracy improves for longer durations, to the point where the error is less than 2.0% for durations greater than 200 ms. The measurements on a Linux system give results similar to those seen when making direct use of cycle counters. This can be seen by comparing the measurements to the Load 1 results in Figure 9.14 (without compensation) and in Figure 9.16 (with compensation). Using compensation, we can achieve better than 0.04% accuracy, even for measurements as long as 300 ms. Thus, gettimeofday performs just as well as directly accessing the cycle counter on this machine. 9.6 Putting it Together: An Experimental Protocol
We can summarize our experimental ﬁndings in the form of a protocol to determine how to answer the ques...
View Full Document
- Spring '10
- The American