See also AT-STYLE TIME SPECIFICATION section in the rrdfetch documentation for other ways to specify time.
ds-name is the name you will use to reference this particular data source from an RRD. A ds-name must be 1 to 19 characters long in the characters [a-zA-Z0-9_].
DST defines the Data Source Type. The remaining arguments of a data source entry depend on the data source type. For GAUGE, COUNTER, DERIVE, and ABSOLUTE the format for a data source entry is:
DS:ds-name:GAUGE | COUNTER | DERIVE | ABSOLUTE:heartbeat:min:max
For COMPUTE data sources, the format is:
In order to decide which data source type to use, review the definitions that follow. Also consult the section on ``HOW TO MEASURE'' for further insight.
NOTE on COUNTER vs DERIVE
by Don Baarda <email@example.com>
If you cannot tolerate ever mistaking the occasional counter reset for a legitimate counter wrap, and would prefer ``Unknowns'' for all legitimate counter wraps and resets, always use DERIVE with min=0. Otherwise, using COUNTER with a suitable max will return correct values for all legitimate counter wraps, mark some counter resets as ``Unknown'', but can mistake some counter resets for a legitimate counter wrap.
For a 5 minute step and 32-bit counter, the probability of mistaking a counter reset for a legitimate wrap is arguably about 0.8% per 1Mbps of maximum bandwidth. Note that this equates to 80% for 100Mbps interfaces, so for high bandwidth interfaces and a 32bit counter, DERIVE with min=0 is probably preferable. If you are using a 64bit counter, just about any max setting will eliminate the possibility of mistaking a reset for a counter wrap.
heartbeat defines the maximum number of seconds that may pass between two updates of this data source before the value of the data source is assumed to be *UNKNOWN*.
min and max define the expected range values for data supplied by a data source. If min and/or max any value outside the defined range will be regarded as *UNKNOWN*. If you do not know or care about min and max, set them to U for unknown. Note that min and max always refer to the processed values of the DS. For a traffic-COUNTER type DS this would be the maximum and minimum data-rate expected from the device.
If information on minimal/maximal expected values is available, always set the min and/or max properties. This will help RRDtool in doing a simple sanity check on the data supplied when running update.
rpn-expression defines the formula used to compute the PDPs of a COMPUTE data source from other data sources in the same <RRD>. It is similar to defining a CDEF argument for the graph command. Please refer to that manual page for a list and description of RPN operations supported. For COMPUTE data sources, the following RPN operations are not supported: COUNT, PREV, TIME, and LTIME. In addition, in defining the RPN expression, the COMPUTE data source may only refer to the names of data source listed previously in the create command. This is similar to the restriction that CDEFs must refer only to DEFs and CDEFs previously defined in the same graph command.
When data is entered into an RRD, it is first fit into time slots of the length defined with the -s option, thus becoming a primary data point.
The data is also processed with the consolidation function (CF) of the archive. There are several consolidation functions that consolidate primary data points via an aggregate function: AVERAGE, MIN, MAX, LAST.
Note that data aggregation inevitably leads to loss of precision and information. The trick is to pick the aggregate function such that the interesting properties of your data is kept across the aggregation process.
The format of RRA line for these consolidation functions is:
RRA:AVERAGE | MIN | MAX | LAST:xff:steps:rows
xff The xfiles factor defines what part of a consolidation interval may be made up from *UNKNOWN* data while the consolidated value is still regarded as known. It is given as the ratio of allowed *UNKNOWN* PDPs to the number of PDPs in the interval. Thus, it ranges from 0 to 1 (exclusive).
steps defines how many of these primary data points are used to build a consolidated data point which then goes into the archive.
rows defines how many generations of data values are kept in an RRA. Obviously, this has to be greater than zero.
These RRAs differ from the true consolidation functions in several ways. First, each of the RRAs is updated once for every primary data point. Second, these RRAs are interdependent. To generate real-time confidence bounds, a matched set of SEASONAL, DEVSEASONAL, DEVPREDICT, and either HWPREDICT or MHWPREDICT must exist. Generating smoothed values of the primary data points requires a SEASONAL RRA and either an HWPREDICT or MHWPREDICT RRA. Aberrant behavior detection requires FAILURES, DEVSEASONAL, SEASONAL, and either HWPREDICT or MHWPREDICT.
The predicted, or smoothed, values are stored in the HWPREDICT or MHWPREDICT RRA. HWPREDICT and MHWPREDICT 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. HWPREDICT adds its seasonal coefficient to the baseline to form a prediction, whereas MHWPREDICT 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; HWPREDICT will predict the seasonality to stay constant as the baseline changes, but MHWPREDICT 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 HWPREDICT, but MHWPREDICT may be substituted in its place.
The predicted deviations are stored in DEVPREDICT (think a standard deviation which can be scaled to yield a confidence band). The FAILURES RRA 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 RRAs to graph confidence bounds and failures appears in rrdgraph.
The SEASONAL and DEVSEASONAL RRAs store the seasonal coefficients for the Holt-Winters forecasting algorithm and the seasonal deviations, respectively. There is one entry per observation time point in the seasonal cycle. For example, if primary data points are generated every five minutes and the seasonal cycle is 1 day, both SEASONAL and DEVSEASONAL will have 288 rows.
In order to simplify the creation for the novice user, in addition to supporting explicit creation of the HWPREDICT, SEASONAL, DEVPREDICT, DEVSEASONAL, and FAILURES RRAs, the RRDtool create command supports implicit creation of the other four when HWPREDICT is specified alone and the final argument rra-num is omitted.
rows specifies the length of the RRA prior to wrap around. Remember that there is a one-to-one correspondence between primary data points and entries in these RRAs. For the HWPREDICT CF, rows should be larger than the seasonal period. If the DEVPREDICT RRA is implicitly created, the default number of rows is the same as the HWPREDICT rows argument. If the FAILURES RRA is implicitly created, rows will be set to the seasonal period argument of the HWPREDICT RRA. Of course, the RRDtool resize command is available if these defaults are not sufficient and the creator wishes to avoid explicit creations of the other specialized function RRAs.
seasonal period specifies the number of primary data points in a seasonal cycle. If SEASONAL and DEVSEASONAL are implicitly created, this argument for those RRAs is set automatically to the value specified by HWPREDICT. If they are explicitly created, the creator should verify that all three seasonal period arguments agree.
alpha is the adaption parameter of the intercept (or baseline) coefficient in the Holt-Winters forecasting algorithm. See rrdtool for a description of this algorithm. alpha must lie between 0 and 1. A value closer to 1 means that more recent observations carry greater weight in predicting the baseline component of the forecast. A value closer to 0 means that past history carries greater weight in predicting the baseline component.
beta is the adaption parameter of the slope (or linear trend) coefficient in the Holt-Winters forecasting algorithm. beta must lie between 0 and 1 and plays the same role as alpha with respect to the predicted linear trend.
gamma is the adaption parameter of the seasonal coefficients in the Holt-Winters forecasting algorithm (HWPREDICT) or the adaption parameter in the exponential smoothing update of the seasonal deviations. It must lie between 0 and 1. If the SEASONAL and DEVSEASONAL RRAs are created implicitly, they will both have the same value for gamma: the value specified for the HWPREDICT alpha argument. Note that because there is one seasonal coefficient (or deviation) for each time point during the seasonal cycle, the adaptation rate is much slower than the baseline. Each seasonal coefficient is only updated (or adapts) when the observed value occurs at the offset in the seasonal cycle corresponding to that coefficient.
If SEASONAL and DEVSEASONAL RRAs are created explicitly, gamma need not be the same for both. Note that gamma can also be changed via the RRDtool tune command.
smoothing-window specifies the fraction of a season that should be averaged around each point. By default, the value of smoothing-window is 0.05, which means each value in SEASONAL and DEVSEASONAL will be occasionally replaced by averaging it with its (seasonal period*0.05) nearest neighbors. Setting smoothing-window to zero will disable the running-average smoother altogether.
rra-num provides the links between related RRAs. If HWPREDICT is specified alone and the other RRAs are created implicitly, then there is no need to worry about this argument. If RRAs are created explicitly, then carefully pay attention to this argument. For each RRA which includes this argument, there is a dependency between that RRA and another RRA. The rra-num argument is the 1-based index in the order of RRA creation (that is, the order they appear in the create command). The dependent RRA for each RRA requiring the rra-num argument is listed here:
threshold is the minimum number of violations (observed values outside the confidence bounds) within a window that constitutes a failure. If the FAILURES RRA is implicitly created, the default value is 7.
window length is the number of time points in the window. Specify an integer greater than or equal to the threshold and less than or equal to 28. The time interval this window represents depends on the interval between primary data points. If the FAILURES RRA is implicitly created, the default value is 9.
RRDtool gets fed samples/updates at arbitrary times. From these it builds Primary Data Points (PDPs) on every ``step'' interval. The PDPs are then accumulated into the RRAs.
The ``heartbeat'' defines the maximum acceptable interval between samples/updates. If the interval between samples is less than ``heartbeat'', then an average rate is calculated and applied for that interval. If the interval between samples is longer than ``heartbeat'', then that entire interval is considered ``unknown''. Note that there are other things that can make a sample interval ``unknown'', such as the rate exceeding limits, or a sample that was explicitly marked as unknown.
The known rates during a PDP's ``step'' interval are used to calculate an average rate for that PDP. If the total ``unknown'' time accounts for more than half the ``step'', the entire PDP is marked as ``unknown''. This means that a mixture of known and ``unknown'' sample times in a single PDP ``step'' may or may not add up to enough ``known'' time to warrent for a known PDP.
The ``heartbeat'' can be short (unusual) or long (typical) relative to the ``step'' interval between PDPs. A short ``heartbeat'' means you require multiple samples per PDP, and if you don't get them mark the PDP unknown. A long heartbeat can span multiple ``steps'', which means it is acceptable to have multiple PDPs calculated from a single sample. An extreme example of this might be a ``step'' of 5 minutes and a ``heartbeat'' of one day, in which case a single sample every day will result in all the PDPs for that entire day period being set to the same average rate. -- Don Baarda <firstname.lastname@example.org>
time| axis| begin__|00| |01| u|02|----* sample1, restart "hb"-timer u|03| / u|04| / u|05| / u|06|/ "hbt" expired u|07| |08|----* sample2, restart "hb" |09| / |10| / u|11|----* sample3, restart "hb" u|12| / u|13| / step1_u|14| / u|15|/ "swt" expired u|16| |17|----* sample4, restart "hb", create "pdp" for step1 = |18| / = unknown due to 10 "u" labled secs > 0.5 * step |19| / |20| / |21|----* sample5, restart "hb" |22| / |23| / |24|----* sample6, restart "hb" |25| / |26| / |27|----* sample7, restart "hb" step2__|28| / |22| / |23|----* sample8, restart "hb", create "pdp" for step1, create "cdp" |24| / |25| /
graphics by email@example.com.
rrdtool create temperature.rrd --step 300 \ DS:temp:GAUGE:600:-273:5000 \ RRA:AVERAGE:0.5:1:1200 \ RRA:MIN:0.5:12:2400 \ RRA:MAX:0.5:12:2400 \ RRA:AVERAGE:0.5:12:2400
This sets up an RRD called temperature.rrd which accepts one temperature value every 300 seconds. If no new data is supplied for more than 600 seconds, the temperature becomes *UNKNOWN*. The minimum acceptable value is -273 and the maximum is 5'000.
A few archive areas are also defined. The first stores the temperatures supplied for 100 hours (1'200 * 300 seconds = 100 hours). The second RRA stores the minimum temperature recorded over every hour (12 * 300 seconds = 1 hour), for 100 days (2'400 hours). The third and the fourth RRA's do the same for the maximum and average temperature, respectively.
rrdtool create monitor.rrd --step 300 \ DS:ifOutOctets:COUNTER:1800:0:4294967295 \ RRA:AVERAGE:0.5:1:2016 \ RRA:HWPREDICT:1440:0.1:0.0035:288
This example is a monitor of a router interface. The first RRA tracks the traffic flow in octets; the second RRA generates the specialized functions RRAs for aberrant behavior detection. Note that the rra-num argument of HWPREDICT is missing, so the other RRAs will implicitly be created with default parameter values. In this example, the forecasting algorithm baseline adapts quickly; in fact the most recent one hour of observations (each at 5 minute intervals) accounts for 75% of the baseline prediction. The linear trend forecast adapts much more slowly. Observations made during the last day (at 288 observations per day) account for only 65% of the predicted linear trend. Note: these computations rely on an exponential smoothing formula described in the LISA 2000 paper.
The seasonal cycle is one day (288 data points at 300 second intervals), and the seasonal adaption parameter will be set to 0.1. The RRD file will store 5 days (1'440 data points) of forecasts and deviation predictions before wrap around. The file will store 1 day (a seasonal cycle) of 0-1 indicators in the FAILURES RRA.
The same RRD file and RRAs are created with the following command, which explicitly creates all specialized function RRAs.
rrdtool create monitor.rrd --step 300 \ DS:ifOutOctets:COUNTER:1800:0:4294967295 \ RRA:AVERAGE:0.5:1:2016 \ RRA:HWPREDICT:1440:0.1:0.0035:288:3 \ RRA:SEASONAL:288:0.1:2 \ RRA:DEVPREDICT:1440:5 \ RRA:DEVSEASONAL:288:0.1:2 \ RRA:FAILURES:288:7:9:5
Of course, explicit creation need not replicate implicit create, a number of arguments could be changed.
rrdtool create proxy.rrd --step 300 \ DS:Total:DERIVE:1800:0:U \ DS:Duration:DERIVE:1800:0:U \ DS:AvgReqDur:COMPUTE:Duration,Requests,0,EQ,1,Requests,IF,/ \ RRA:AVERAGE:0.5:1:2016
This example is monitoring the average request duration during each 300 sec interval for requests processed by a web proxy during the interval. In this case, the proxy exposes two counters, the number of requests processed since boot and the total cumulative duration of all processed requests. Clearly these counters both have some rollover point, but using the DERIVE data source also handles the reset that occurs when the web proxy is stopped and restarted.
In the RRD, the first data source stores the requests per second rate during the interval. The second data source stores the total duration of all requests processed during the interval divided by 300. The COMPUTE data source divides each PDP of the AccumDuration by the corresponding PDP of TotalRequests and stores the average request duration. The remainder of the RPN expression handles the divide by zero case.