diff --git a/doc/rrdcreate.1 b/doc/rrdcreate.1
index b18228fe834564cb1fde37d5e2db75838a886f81..53fc8d98c073cdbf48cd0d603e26bdf77ab3ed10 100644 (file)
--- a/doc/rrdcreate.1
+++ b/doc/rrdcreate.1
.\" ========================================================================
.\"
.IX Title "RRDCREATE 1"
-.TH RRDCREATE 1 "2007-11-20" "1.2.26" "rrdtool"
+.TH RRDCREATE 1 "2008-05-12" "1.3rc9" "rrdtool"
.SH "NAME"
rrdcreate \- Set up a new Round Robin Database
.SH "SYNOPSIS"
The data is also processed with the consolidation function (\fI\s-1CF\s0\fR) of
the archive. There are several consolidation functions that
consolidate primary data points via an aggregate function: \fB\s-1AVERAGE\s0\fR,
-\&\fB\s-1MIN\s0\fR, \fB\s-1MAX\s0\fR, \fB\s-1LAST\s0\fR. The format of \fB\s-1RRA\s0\fR line for these
+\&\fB\s-1MIN\s0\fR, \fB\s-1MAX\s0\fR, \fB\s-1LAST\s0\fR.
+.RS 8
+.IP "\s-1AVERAGE\s0" 4
+.IX Item "AVERAGE"
+the average of the data points is stored.
+.IP "\s-1MIN\s0" 4
+.IX Item "MIN"
+the smallest of the data points is stored.
+.IP "\s-1MAX\s0" 4
+.IX Item "MAX"
+the largest of the data points is stored.
+.IP "\s-1LAST\s0" 4
+.IX Item "LAST"
+the last data points is used.
+.RE
+.RS 8
+.Sp
+Note that data aggregation inevitably leads to loss of precision and
+information. The trick is to pick the aggregate function such that the
+\&\fIinteresting\fR properties of your data is kept across the aggregation
+process.
+.Sp
+The format of \fB\s-1RRA\s0\fR line for these
consolidation functions is:
.Sp
\&\fB\s-1RRA:\s0\fR\fI\s-1AVERAGE\s0 | \s-1MIN\s0 | \s-1MAX\s0 | \s-1LAST\s0\fR\fB:\fR\fIxff\fR\fB:\fR\fIsteps\fR\fB:\fR\fIrows\fR
a \fIconsolidated data point\fR which then goes into the archive.
.Sp
\&\fIrows\fR defines how many generations of data values are kept in an \fB\s-1RRA\s0\fR.
+.RE
.SH "Aberrant Behavior Detection with Holt-Winters Forecasting"
.IX Header "Aberrant Behavior Detection with Holt-Winters Forecasting"
In addition to the aggregate functions, there are a set of specialized
.IP "\(bu" 4
\&\fB\s-1RRA:\s0\fR\fI\s-1HWPREDICT\s0\fR\fB:\fR\fIrows\fR\fB:\fR\fIalpha\fR\fB:\fR\fIbeta\fR\fB:\fR\fIseasonal period\fR[\fB:\fR\fIrra-num\fR]
.IP "\(bu" 4
-\&\fB\s-1RRA:\s0\fR\fI\s-1SEASONAL\s0\fR\fB:\fR\fIseasonal period\fR\fB:\fR\fIgamma\fR\fB:\fR\fIrra-num\fR
+\&\fB\s-1RRA:\s0\fR\fI\s-1MHWPREDICT\s0\fR\fB:\fR\fIrows\fR\fB:\fR\fIalpha\fR\fB:\fR\fIbeta\fR\fB:\fR\fIseasonal period\fR[\fB:\fR\fIrra-num\fR]
+.IP "\(bu" 4
+\&\fB\s-1RRA:\s0\fR\fI\s-1SEASONAL\s0\fR\fB:\fR\fIseasonal period\fR\fB:\fR\fIgamma\fR\fB:\fR\fIrra-num\fR[\fB:smoothing\-window=\fR\fIfraction\fR]
.IP "\(bu" 4
-\&\fB\s-1RRA:\s0\fR\fI\s-1DEVSEASONAL\s0\fR\fB:\fR\fIseasonal period\fR\fB:\fR\fIgamma\fR\fB:\fR\fIrra-num\fR
+\&\fB\s-1RRA:\s0\fR\fI\s-1DEVSEASONAL\s0\fR\fB:\fR\fIseasonal period\fR\fB:\fR\fIgamma\fR\fB:\fR\fIrra-num\fR[\fB:smoothing\-window=\fR\fIfraction\fR]
.IP "\(bu" 4
\&\fB\s-1RRA:\s0\fR\fI\s-1DEVPREDICT\s0\fR\fB:\fR\fIrows\fR\fB:\fR\fIrra-num\fR
.IP "\(bu" 4
These \fBRRAs\fR differ from the true consolidation functions in several ways.
First, each of the \fB\s-1RRA\s0\fRs is updated once for every primary data point.
Second, these \fBRRAs\fR are interdependent. To generate real-time confidence
-bounds, a matched set of \s-1HWPREDICT\s0, \s-1SEASONAL\s0, \s-1DEVSEASONAL\s0, and
-\&\s-1DEVPREDICT\s0 must exist. Generating smoothed values of the primary data points
-requires both a \s-1HWPREDICT\s0 \fB\s-1RRA\s0\fR and \s-1SEASONAL\s0 \fB\s-1RRA\s0\fR. Aberrant behavior
-detection requires \s-1FAILURES\s0, \s-1HWPREDICT\s0, \s-1DEVSEASONAL\s0, and \s-1SEASONAL\s0.
-.PP
-The actual predicted, or smoothed, values are stored in the \s-1HWPREDICT\s0
-\&\fB\s-1RRA\s0\fR. The predicted deviations are stored in \s-1DEVPREDICT\s0 (think a standard
-deviation which can be scaled to yield a confidence band). The \s-1FAILURES\s0
-\&\fB\s-1RRA\s0\fR stores binary indicators. A 1 marks the indexed observation as
-failure; that is, the number of confidence bounds violations in the
-preceding window of observations met or exceeded a specified threshold. An
-example of using these \fBRRAs\fR to graph confidence bounds and failures
-appears in rrdgraph.
+bounds, a matched set of \s-1SEASONAL\s0, \s-1DEVSEASONAL\s0, \s-1DEVPREDICT\s0, and either
+\&\s-1HWPREDICT\s0 or \s-1MHWPREDICT\s0 must exist. Generating smoothed values of the primary
+data points requires a \s-1SEASONAL\s0 \fB\s-1RRA\s0\fR and either an \s-1HWPREDICT\s0 or \s-1MHWPREDICT\s0
+\&\fB\s-1RRA\s0\fR. Aberrant behavior detection requires \s-1FAILURES\s0, \s-1DEVSEASONAL\s0, \s-1SEASONAL\s0,
+and either \s-1HWPREDICT\s0 or \s-1MHWPREDICT\s0.
+.PP
+The predicted, or smoothed, values are stored in the \s-1HWPREDICT\s0 or \s-1MHWPREDICT\s0
+\&\fB\s-1RRA\s0\fR. \s-1HWPREDICT\s0 and \s-1MHWPREDICT\s0 are actually two variations on the
+Holt-Winters method. They are interchangeable. Both attempt to decompose data
+into three components: a baseline, a trend, and a seasonal coefficient.
+\&\s-1HWPREDICT\s0 adds its seasonal coefficient to the baseline to form a prediction, whereas
+\&\s-1MHWPREDICT\s0 multiplies its seasonal coefficient by the baseline to form a
+prediction. The difference is noticeable when the baseline changes
+significantly in the course of a season; \s-1HWPREDICT\s0 will predict the seasonality
+to stay constant as the baseline changes, but \s-1MHWPREDICT\s0 will predict the
+seasonality to grow or shrink in proportion to the baseline. The proper choice
+of method depends on the thing being modeled. For simplicity, the rest of this
+discussion will refer to \s-1HWPREDICT\s0, but \s-1MHWPREDICT\s0 may be substituted in its
+place.
+.PP
+The predicted deviations are stored in \s-1DEVPREDICT\s0 (think a standard deviation
+which can be scaled to yield a confidence band). The \s-1FAILURES\s0 \fB\s-1RRA\s0\fR stores
+binary indicators. A 1 marks the indexed observation as failure; that is, the
+number of confidence bounds violations in the preceding window of observations
+met or exceeded a specified threshold. An example of using these \fBRRAs\fR to graph
+confidence bounds and failures appears in rrdgraph.
.PP
The \s-1SEASONAL\s0 and \s-1DEVSEASONAL\s0 \fBRRAs\fR store the seasonal coefficients for the
Holt-Winters forecasting algorithm and the seasonal deviations, respectively.
@@ -394,6 +432,13 @@ If \s-1SEASONAL\s0 and \s-1DEVSEASONAL\s0 \fBRRAs\fR are created explicitly, \fI
be the same for both. Note that \fIgamma\fR can also be changed via the
\&\fBRRDtool\fR \fItune\fR command.
.PP
+\&\fIsmoothing-window\fR specifies the fraction of a season that should be
+averaged around each point. By default, the value of \fIsmoothing-window\fR is
+0.05, which means each value in \s-1SEASONAL\s0 and \s-1DEVSEASONAL\s0 will be occasionally
+replaced by averaging it with its (\fIseasonal period\fR*0.05) nearest neighbors.
+Setting \fIsmoothing-window\fR to zero will disable the running-average smoother
+altogether.
+.PP
\&\fIrra-num\fR provides the links between related \fBRRAs\fR. If \s-1HWPREDICT\s0 is
specified alone and the other \fBRRAs\fR are created implicitly, then
there is no need to worry about this argument. If \fBRRAs\fR are created
It may help you to sort out why all this *UNKNOWN* data is popping
up in your databases:
.PP
-RRDtool gets fed samples at arbitrary times. From these it builds Primary
-Data Points (PDPs) at exact times on every \*(L"step\*(R" interval. The PDPs are
-then accumulated into RRAs.
+RRDtool gets fed samples/updates at arbitrary times. From these it builds Primary
+Data Points (PDPs) on every \*(L"step\*(R" interval. The PDPs are
+then accumulated into the RRAs.
.PP
The \*(L"heartbeat\*(R" defines the maximum acceptable interval between
-samples. If the interval between samples is less than \*(L"heartbeat\*(R",
+samples/updates. If the interval between samples is less than \*(L"heartbeat\*(R",
then an average rate is calculated and applied for that interval. If
the interval between samples is longer than \*(L"heartbeat\*(R", then that
entire interval is considered \*(L"unknown\*(R". Note that there are other
things that can make a sample interval \*(L"unknown\*(R", such as the rate
-exceeding limits, or even an \*(L"unknown\*(R" input sample.
+exceeding limits, or a sample that was explicitly marked as unknown.
.PP
The known rates during a \s-1PDP\s0's \*(L"step\*(R" interval are used to calculate
-an average rate for that \s-1PDP\s0. Also, if the total \*(L"unknown\*(R" time during
-the \*(L"step\*(R" interval exceeds the \*(L"heartbeat\*(R", the entire \s-1PDP\s0 is marked
+an average rate for that \s-1PDP\s0. If the total \*(L"unknown\*(R" time accounts for
+more than \fBhalf\fR the \*(L"step\*(R", the entire \s-1PDP\s0 is marked
as \*(L"unknown\*(R". This means that a mixture of known and \*(L"unknown\*(R" sample
-times in a single \s-1PDP\s0 \*(L"step\*(R" may or may not add up to enough \*(L"unknown\*(R"
-time to exceed \*(L"heartbeat\*(R" and hence mark the whole \s-1PDP\s0 \*(L"unknown\*(R". So
-\&\*(L"heartbeat\*(R" is not only the maximum acceptable interval between
-samples, but also the maximum acceptable amount of \*(L"unknown\*(R" time per
-\&\s-1PDP\s0 (obviously this is only significant if you have \*(L"heartbeat\*(R" less
-than \*(L"step\*(R").
+times in a single \s-1PDP\s0 \*(L"step\*(R" may or may not add up to enough \*(L"known\*(R"
+time to warrent for a known \s-1PDP\s0.
.PP
The \*(L"heartbeat\*(R" can be short (unusual) or long (typical) relative to
the \*(L"step\*(R" interval between PDPs. A short \*(L"heartbeat\*(R" means you
\& u|15|/ "swt" expired
\& u|16|
\& |17|\-\-\-\-* sample4, restart "hb", create "pdp" for step1 =
-\& |18| / = unknown due to 10 "u" labled secs > "hb"
+\& |18| / = unknown due to 10 "u" labled secs > 0.5 * step
\& |19| /
\& |20| /
\& |21|\-\-\-\-* sample5, restart "hb"