06/01/1993 09:00 AM

Computer Engineering

This thesis is a study of the failure statistics of hosts on the Internet. To cross-verify our estimates of failure rates, we measure three different kinds of observations about hosts on the Internet - the time interval between failure, the number of failures in a certain time interval, and the length of time a host has been up at an arbitrary instant of time. These observations are made on different samples of hosts on the Internet. We present statistical results that allow us to derive estimates about the failure statistics of the population from these three different types of observations. The estimates from the three experiments are slightly different but comparable, providing three independent sources of evidence about the failure statistics of hosts on the Internet. We also attempt to explain the observations by modeling the statistical process with two distributions - the exponential and Weibull distribution. The observed data are explained well (in a statistical sense) by both models, but the Weibull model fits the observed sample better. We believe this to be evidence that the failure rates of hosts on the Internet are nearly constant, but with a high initial value that quickly decreases to a constant as the machine continues to run. Notes: Masters Thesis