Algorithm for Detecting Significant Changes in Performance Metric Values
Original Publication Date: 1995-Jul-01
Included in the Prior Art Database: 2005-Mar-30
Disclosed is a technique for detecting significant changes in key system performance metrics.
Algorithm for Detecting Significant Changes in
a technique for detecting significant changes in
key system performance metrics.
values for a key performance metric are collected
over time and recorded in a historical repository. Also, data in the
repository is periodically examined for evidence of system
performance problems. Certain types of system performance problems
are suggested by sudden, dramatic changes in key metrics - therefore,
automatically detecting such changes is an important component of a
performance problem detection application.
Gather the most recent N measurement values of a
particular metric X;
designate these as X(1), X(2), ... X(N). For example, suppose
metric X indicates the number of Zombie processes associated with a
particular userid in a UNIX system; values for X are collected
periodically, say once every second. X(N) is the most recent
measurement value, X(N-1) is the measurement value collected one
second ago, X(N-2) is the measurement value collected two seconds
ago, and so on.
The X(1..N) measurements are collected under the same
(e.g., same time of day), and so it is reasonable to assume that they
represent random samples from the same distribution.
Compute the mean and second moment for values
the hypothesis that X(N) is a sample from a Normal distribution
having that same mean and second moment. If this hypot...