=head1 DESCRIPTION
If you have ever used a traditional HP calculator you already know
-B<RPN>. The idea behind B<RPN> is that you have a stack and push
+B<RPN> (Reverse Polish Notation).
+The idea behind B<RPN> is that you have a stack and push
your data onto this stack. Whenever you execute an operation, it
takes as many elements from the stack as needed. Pushing is done
implicitly, so whenever you specify a number or a variable, it gets
pushed onto the stack automatically.
-At the end of the calculation there should be one and only one
-value left on the stack. This is the outcome of the function and
-this is what is put into the I<vname>. For B<CDEF> instructions,
-the stack is processed for each data point on the graph. B<VDEF>
-instructions work on an entire data set in one run.
+At the end of the calculation there should be one and only one value left on
+the stack. This is the outcome of the function and this is what is put into
+the I<vname>. For B<CDEF> instructions, the stack is processed for each
+data point on the graph. B<VDEF> instructions work on an entire data set in
+one run. Note, that currently B<VDEF> instructions only support a limited
+list of functions.
Example: C<VDEF:maximum=mydata,MAXIMUM>
Example: C<CDEF:mydatabits=mydata,8,*>
This means: push variable I<mydata>, push the number 8, execute
-the operator I<+>. The operator needs two elements and uses those
+the operator I<*>. The operator needs two elements and uses those
to return one value. This value is then stored in I<mydatabits>.
As you may have guessed, this instruction means nothing more than
I<mydatabits = mydata * 8>. The real power of B<RPN> lies in the
Pop two elements from the stack, compare them for the selected condition
and return 1 for true or 0 for false. Comparing an I<unknown> or an
-I<infinite> value will always result in 0 (false).
+I<infinite> value will result in I<unknown> returned ... which will also be
+treated as false by the B<IF> call.
B<UN, ISINF>
Add, subtract, multiply, divide, modulo
+B<ADDNAN>
+
+NAN-safe addition. If one parameter is NAN/UNKNOWN it'll be treated as
+zero. If both parameters are NAN/UNKNOWN, NAN/UNKNOWN will be returned.
+
B<SIN, COS, LOG, EXP, SQRT>
Sine and cosine (input in radians), log and exp (natural logarithm),
Arctangent (output in radians).
+B<ATAN2>
+
+Arctangent of y,x components (output in radians).
+This pops one element from the stack, the x (cosine) component, and then
+a second, which is the y (sine) component.
+It then pushes the arctangent of their ratio, resolving the ambiguity between
+quadrants.
+
+Example: C<CDEF:angle=Y,X,ATAN2,RAD2DEG> will convert C<X,Y>
+components into an angle in degrees.
+
B<FLOOR, CEIL>
Round down or up to the nearest integer.
-Z<>
+B<DEG2RAD, RAD2DEG>
+
+Convert angle in degrees to radians, or radians to degrees.
+
+B<ABS>
+
+Take the absolute value.
=item Set Operations
compute the average of the values v1 to v6 after removing the smallest and
largest.
-B<TREND>
+B<AVG>
+
+Pop one element (I<count>) from the stack. Now pop I<count> elements and build the
+average, ignoring all UNKNOWN values in the process.
+
+Example: C<CDEF:x=a,b,c,d,4,AVG>
+
+B<TREND, TRENDNAN>
Create a "sliding window" average of another data series.
Value at sample (t1) will be the average between (t1-delay) and (t1)
Value at sample (t2) will be the average between (t2-delay) and (t2)
+TRENDNAN is - in contrast to TREND - NAN-safe. If you use TREND and one
+source value is NAN the complete sliding window is affected. The TRENDNAN
+operation ignores all NAN-values in a sliding window and computes the
+average of the remaining values.
+
+B<PREDICT, PREDICTSIGMA>
+
+Create a "sliding window" average/sigma of another data series, that also
+shifts the data series by given amounts of of time as well
+
+Usage - explicit stating shifts:
+CDEF:predict=<shift n>,...,<shift 1>,n,<window>,x,PREDICT
+CDEF:sigma=<shift n>,...,<shift 1>,n,<window>,x,PREDICTSIGMA
+
+Usage - shifts defined as a base shift and a number of time this is applied
+CDEF:predict=<shift multiplier>,-n,<window>,x,PREDICT
+CDEF:sigma=<shift multiplier>,-n,<window>,x,PREDICTSIGMA
+
+Example:
+CDEF:predict=172800,86400,2,1800,x,PREDICT
+
+This will create a half-hour (1800 second) sliding window average/sigma of x, that
+average is essentially computed as shown here:
+
+ +---!---!---!---!---!---!---!---!---!---!---!---!---!---!---!---!---!--->
+ now
+ shift 1 t0
+ <----------------------->
+ window
+ <--------------->
+ shift 2
+ <----------------------------------------------->
+ window
+ <--------------->
+ shift 1 t1
+ <----------------------->
+ window
+ <--------------->
+ shift 2
+ <----------------------------------------------->
+ window
+ <--------------->
+
+ Value at sample (t0) will be the average between (t0-shift1-window) and (t0-shift1)
+ and between (t0-shift2-window) and (t0-shift2)
+ Value at sample (t1) will be the average between (t1-shift1-window) and (t1-shift1)
+ and between (t1-shift2-window) and (t1-shift2)
+
+
+The function is by design NAN-safe.
+This also allows for extrapolation into the future (say a few days)
+- you may need to define the data series whit the optional start= parameter, so that
+the source data series has enough data to provide prediction also at the beginning of a graph...
+
+Here an example, that will create a 10 day graph that also shows the
+prediction 3 days into the future with its uncertainty value (as defined by avg+-4*sigma)
+This also shows if the prediction is exceeded at a certain point.
+
+rrdtool graph image.png --imgformat=PNG \
+ --start=-7days --end=+3days --width=1000 --height=200 --alt-autoscale-max \
+ DEF:value=value.rrd:value:AVERAGE:start=-14days \
+ LINE1:value#ff0000:value \
+ CDEF:predict=86400,-7,1800,value,PREDICT \
+ CDEF:sigma=86400,-7,1800,value,PREDICTSIGMA \
+ CDEF:upper=predict,sigma,3,*,+ \
+ CDEF:lower=predict,sigma,3,*,- \
+ LINE1:predict#00ff00:prediction \
+ LINE1:upper#0000ff:upper\ certainty\ limit \
+ LINE1:lower#0000ff:lower\ certainty\ limit \
+ CDEF:exceeds=value,UN,0,value,lower,upper,LIMIT,UN,IF \
+ TICK:exceeds#aa000080:1
+
+Note: Experience has shown that a factor between 3 and 5 to scale sigma is a good
+discriminator to detect abnormal behavior. This obviously depends also on the type
+of data and how "noisy" the data series is.
+
+This prediction can only be used for short term extrapolations - say a few days into the future-
+
=item Special values
B<UNKN>
you to make calculations based on the position of the value within
the data set. This function cannot be used in B<VDEF> instructions.
-Z<>
-
=item Time
Time inside RRDtool is measured in seconds since the epoch. The
=head1 VARIABLES
-These operators work only on B<VDEF> statements.
+These operators work only on B<VDEF> statements. Note that currently ONLY these work for B<VDEF>.
=over 4
Example: C<VDEF:avg=mydata,AVERAGE>
+=item STDEV
+
+Returns the standard deviation of the values.
+
+Example: C<VDEF:stdev=mydata,STDEV>
+
=item LAST, FIRST
Return the last/first value including its time. The time for
=item TOTAL
Returns the rate from each defined time slot multiplied with the
-step size. This can, for instance, return total bytes transfered
+step size. This can, for instance, return total bytes transferred
when you have logged bytes per second. The time component returns
the number of seconds.
Example: C<VDEF:total=mydata,TOTAL>
-=item PERCENT
+=item PERCENT, PERCENTNAN
This should follow a B<DEF> or B<CDEF> I<vname>. The I<vname> is popped,
another number is popped which is a certain percentage (0..100). The
data set is then sorted and the value returned is chosen such that
I<percentage> percent of the values is lower or equal than the result.
+For PERCENTNAN I<Unknown> values are ignored, but for PERCENT
I<Unknown> values are considered lower than any finite number for this
purpose so if this operator returns an I<unknown> you have quite a lot
of them in your data. B<Inf>inite numbers are lesser, or more, than the
(NaN E<lt> -INF E<lt> finite values E<lt> INF)
Example: C<VDEF:perc95=mydata,95,PERCENT>
+ C<VDEF:percnan95=mydata,95,PERCENTNAN>
+
+=item LSLSLOPE, LSLINT, LSLCORREL
+
+Return the parameters for a B<L>east B<S>quares B<L>ine I<(y = mx +b)>
+which approximate the provided dataset. LSLSLOPE is the slope I<(m)> of
+the line related to the COUNT position of the data. LSLINT is the
+y-intercept I<(b)>, which happens also to be the first data point on the
+graph. LSLCORREL is the Correlation Coefficient (also know as Pearson's
+Product Moment Correlation Coefficient). It will range from 0 to +/-1
+and represents the quality of fit for the approximation.
+
+Example: C<VDEF:slope=mydata,LSLSLOPE>
=back
=head1 AUTHOR
-Program by Tobias Oetiker E<lt>oetiker@ee.ethz.chE<gt>
+Program by Tobias Oetiker E<lt>tobi@oetiker.chE<gt>
-This manual page by Alex van den Bogaerdt E<lt>alex@ergens.op.het.netE<gt>
+This manual page by Alex van den Bogaerdt E<lt>alex@vandenbogaerdt.nlE<gt>
+with corrections and/or additions by several people
projects / rrdtool.git / blobdiff