Code

Imported upstream SVN snapshot 1.4~rc2+20090928.
[pkg-rrdtool.git] / doc / rrdcreate.html
1 <?xml version="1.0" ?>
2 <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd">
3 <html xmlns="http://www.w3.org/1999/xhtml">
4 <head>
5 <title>rrdcreate</title>
6 <meta http-equiv="content-type" content="text/html; charset=utf-8" />
7 <link rev="made" href="mailto:root@localhost" />
8 </head>
10 <body style="background-color: white">
13 <!-- INDEX BEGIN -->
14 <div name="index">
15 <p><a name="__index__"></a></p>
16 <!--
18 <ul>
20         <li><a href="#name">NAME</a></li>
21         <li><a href="#synopsis">SYNOPSIS</a></li>
22         <li><a href="#description">DESCRIPTION</a></li>
23         <ul>
25                 <li><a href="#filename"><em>filename</em></a></li>
26                 <li><a href="#__start__b_start_time__default__now___10s_"><strong>-start</strong>|<strong>-b</strong> <em>start time</em> (default: now - 10s)</a></li>
27                 <li><a href="#__step__s_step__default__300_seconds_"><strong>-step</strong>|<strong>-s</strong> <em>step</em> (default: 300 seconds)</a></li>
28                 <li><a href="#ds_ds_name_dst_dst_arguments"><strong>DS:</strong><em>ds-name</em><strong>:</strong><em>DST</em><strong>:</strong><em>dst arguments</em></a></li>
29                 <li><a href="#rra_cf_cf_arguments"><strong>RRA:</strong><em>CF</em><strong>:</strong><em>cf arguments</em></a></li>
30         </ul>
32         <li><a href="#aberrant_behavior_detection_with_holt_winters_forecasting">Aberrant Behavior Detection with Holt-Winters Forecasting</a></li>
33         <li><a href="#the_heartbeat_and_the_step">The HEARTBEAT and the STEP</a></li>
34         <li><a href="#how_to_measure">HOW TO MEASURE</a></li>
35         <li><a href="#example">EXAMPLE</a></li>
36         <li><a href="#example_2">EXAMPLE 2</a></li>
37         <li><a href="#example_3">EXAMPLE 3</a></li>
38         <li><a href="#author">AUTHOR</a></li>
39 </ul>
41 -->
44 </div>
45 <!-- INDEX END -->
47 <p>
48 </p>
49 <h1><a name="name">NAME</a></h1>
50 <p>rrdcreate - Set up a new Round Robin Database</p>
51 <p>
52 </p>
53 <hr />
54 <h1><a name="synopsis">SYNOPSIS</a></h1>
55 <p><strong>rrdtool</strong> <strong>create</strong> <em>filename</em>
56 [<strong>--start</strong>|<strong>-b</strong>&nbsp;<em>start&nbsp;time</em>]
57 [<strong>--step</strong>|<strong>-s</strong>&nbsp;<em>step</em>]
58 [<strong>DS:</strong><em>ds-name</em><strong>:</strong><em>DST</em><strong>:</strong><em>dst&nbsp;arguments</em>]
59 [<strong>RRA:</strong><em>CF</em><strong>:</strong><em>cf&nbsp;arguments</em>]</p>
60 <p>
61 </p>
62 <hr />
63 <h1><a name="description">DESCRIPTION</a></h1>
64 <p>The create function of RRDtool lets you set up new Round Robin
65 Database (<strong>RRD</strong>) files.  The file is created at its final, full size
66 and filled with <em>*UNKNOWN*</em> data.</p>
67 <p>
68 </p>
69 <h2><a name="filename"><em>filename</em></a></h2>
70 <p>The name of the <strong>RRD</strong> you want to create. <strong>RRD</strong> files should end
71 with the extension <em class="file">.rrd</em>. However, <strong>RRDtool</strong> will accept any
72 filename.</p>
73 <p>
74 </p>
75 <h2><a name="__start__b_start_time__default__now___10s_"><strong>--start</strong>|<strong>-b</strong> <em>start time</em> (default: now - 10s)</a></h2>
76 <p>Specifies the time in seconds since 1970-01-01 UTC when the first
77 value should be added to the <strong>RRD</strong>. <strong>RRDtool</strong> will not accept
78 any data timed before or at the time specified.</p>
79 <p>See also AT-STYLE TIME SPECIFICATION section in the
80 <em>rrdfetch</em> documentation for other ways to specify time.</p>
81 <p>
82 </p>
83 <h2><a name="__step__s_step__default__300_seconds_"><strong>--step</strong>|<strong>-s</strong> <em>step</em> (default: 300 seconds)</a></h2>
84 <p>Specifies the base interval in seconds with which data will be fed
85 into the <strong>RRD</strong>.</p>
86 <p>
87 </p>
88 <h2><a name="ds_ds_name_dst_dst_arguments"><strong>DS:</strong><em>ds-name</em><strong>:</strong><em>DST</em><strong>:</strong><em>dst arguments</em></a></h2>
89 <p>A single <strong>RRD</strong> can accept input from several data sources (<strong>DS</strong>),
90 for example incoming and outgoing traffic on a specific communication
91 line. With the <strong>DS</strong> configuration option you must define some basic
92 properties of each data source you want to store in the <strong>RRD</strong>.</p>
93 <p><em>ds-name</em> is the name you will use to reference this particular data
94 source from an <strong>RRD</strong>. A <em>ds-name</em> must be 1 to 19 characters long in
95 the characters [a-zA-Z0-9_].</p>
96 <p><em>DST</em> defines the Data Source Type. The remaining arguments of a
97 data source entry depend on the data source type. For GAUGE, COUNTER,
98 DERIVE, and ABSOLUTE the format for a data source entry is:</p>
99 <p><strong>DS:</strong><em>ds-name</em><strong>:</strong><em>GAUGE | COUNTER | DERIVE | ABSOLUTE</em><strong>:</strong><em>heartbeat</em><strong>:</strong><em>min</em><strong>:</strong><em>max</em></p>
100 <p>For COMPUTE data sources, the format is:</p>
101 <p><strong>DS:</strong><em>ds-name</em><strong>:</strong><em>COMPUTE</em><strong>:</strong><em>rpn-expression</em></p>
102 <p>In order to decide which data source type to use, review the
103 definitions that follow. Also consult the section on &quot;HOW TO MEASURE&quot;
104 for further insight.</p>
105 <dl>
106 <dt><strong><a name="gauge" class="item"><strong>GAUGE</strong></a></strong></dt>
108 <dd>
109 <p>is for things like temperatures or number of people in a room or the
110 value of a RedHat share.</p>
111 </dd>
112 <dt><strong><a name="counter" class="item"><strong>COUNTER</strong></a></strong></dt>
114 <dd>
115 <p>is for continuous incrementing counters like the ifInOctets counter in
116 a router. The <strong>COUNTER</strong> data source assumes that the counter never
117 decreases, except when a counter overflows.  The update function takes
118 the overflow into account.  The counter is stored as a per-second
119 rate. When the counter overflows, RRDtool checks if the overflow
120 happened at the 32bit or 64bit border and acts accordingly by adding
121 an appropriate value to the result.</p>
122 </dd>
123 <dt><strong><a name="derive" class="item"><strong>DERIVE</strong></a></strong></dt>
125 <dd>
126 <p>will store the derivative of the line going from the last to the
127 current value of the data source. This can be useful for gauges, for
128 example, to measure the rate of people entering or leaving a
129 room. Internally, derive works exactly like COUNTER but without
130 overflow checks. So if your counter does not reset at 32 or 64 bit you
131 might want to use DERIVE and combine it with a MIN value of 0.</p>
132 <p><strong>NOTE on COUNTER vs DERIVE</strong></p>
133 <p>by Don Baarda &lt;<a href="mailto:don.baarda@baesystems.com">don.baarda@baesystems.com</a>&gt;</p>
134 <p>If you cannot tolerate ever mistaking the occasional counter reset for a
135 legitimate counter wrap, and would prefer &quot;Unknowns&quot; for all legitimate
136 counter wraps and resets, always use DERIVE with min=0. Otherwise, using
137 COUNTER with a suitable max will return correct values for all legitimate
138 counter wraps, mark some counter resets as &quot;Unknown&quot;, but can mistake some
139 counter resets for a legitimate counter wrap.</p>
140 <p>For a 5 minute step and 32-bit counter, the probability of mistaking a
141 counter reset for a legitimate wrap is arguably about 0.8% per 1Mbps of
142 maximum bandwidth. Note that this equates to 80% for 100Mbps interfaces, so
143 for high bandwidth interfaces and a 32bit counter, DERIVE with min=0 is
144 probably preferable. If you are using a 64bit counter, just about any max
145 setting will eliminate the possibility of mistaking a reset for a counter
146 wrap.</p>
147 </dd>
148 <dt><strong><a name="absolute" class="item"><strong>ABSOLUTE</strong></a></strong></dt>
150 <dd>
151 <p>is for counters which get reset upon reading. This is used for fast counters
152 which tend to overflow. So instead of reading them normally you reset them
153 after every read to make sure you have a maximum time available before the
154 next overflow. Another usage is for things you count like number of messages
155 since the last update.</p>
156 </dd>
157 <dt><strong><a name="compute" class="item"><strong>COMPUTE</strong></a></strong></dt>
159 <dd>
160 <p>is for storing the result of a formula applied to other data sources
161 in the <strong>RRD</strong>. This data source is not supplied a value on update, but
162 rather its Primary Data Points (PDPs) are computed from the PDPs of
163 the data sources according to the rpn-expression that defines the
164 formula. Consolidation functions are then applied normally to the PDPs
165 of the COMPUTE data source (that is the rpn-expression is only applied
166 to generate PDPs). In database software, such data sets are referred
167 to as &quot;virtual&quot; or &quot;computed&quot; columns.</p>
168 </dd>
169 </dl>
170 <p><em>heartbeat</em> defines the maximum number of seconds that may pass
171 between two updates of this data source before the value of the
172 data source is assumed to be <em>*UNKNOWN*</em>.</p>
173 <p><em>min</em> and <em>max</em> define the expected range values for data supplied by a
174 data source. If <em>min</em> and/or <em>max</em> any value outside the defined range
175 will be regarded as <em>*UNKNOWN*</em>. If you do not know or care about min and
176 max, set them to U for unknown. Note that min and max always refer to the
177 processed values of the DS. For a traffic-<strong>COUNTER</strong> type DS this would be
178 the maximum and minimum data-rate expected from the device.</p>
179 <p><em>If information on minimal/maximal expected values is available,
180 always set the min and/or max properties. This will help RRDtool in
181 doing a simple sanity check on the data supplied when running update.</em></p>
182 <p><em>rpn-expression</em> defines the formula used to compute the PDPs of a
183 COMPUTE data source from other data sources in the same &lt;RRD&gt;. It is
184 similar to defining a <strong>CDEF</strong> argument for the graph command. Please
185 refer to that manual page for a list and description of RPN operations
186 supported. For COMPUTE data sources, the following RPN operations are
187 not supported: COUNT, PREV, TIME, and LTIME. In addition, in defining
188 the RPN expression, the COMPUTE data source may only refer to the
189 names of data source listed previously in the create command. This is
190 similar to the restriction that <strong>CDEF</strong>s must refer only to <strong>DEF</strong>s
191 and <strong>CDEF</strong>s previously defined in the same graph command.</p>
192 <p>
193 </p>
194 <h2><a name="rra_cf_cf_arguments"><strong>RRA:</strong><em>CF</em><strong>:</strong><em>cf arguments</em></a></h2>
195 <p>The purpose of an <strong>RRD</strong> is to store data in the round robin archives
196 (<strong>RRA</strong>). An archive consists of a number of data values or statistics for
197 each of the defined data-sources (<strong>DS</strong>) and is defined with an <strong>RRA</strong> line.</p>
198 <p>When data is entered into an <strong>RRD</strong>, it is first fit into time slots
199 of the length defined with the <strong>-s</strong> option, thus becoming a <em>primary
200 data point</em>.</p>
201 <p>The data is also processed with the consolidation function (<em>CF</em>) of
202 the archive. There are several consolidation functions that
203 consolidate primary data points via an aggregate function: <strong>AVERAGE</strong>,
204 <strong>MIN</strong>, <strong>MAX</strong>, <strong>LAST</strong>.</p>
205 <dl>
206 <dt><strong><a name="average" class="item">AVERAGE</a></strong></dt>
208 <dd>
209 <p>the average of the data points is stored.</p>
210 </dd>
211 <dt><strong><a name="min" class="item">MIN</a></strong></dt>
213 <dd>
214 <p>the smallest of the data points is stored.</p>
215 </dd>
216 <dt><strong><a name="max" class="item">MAX</a></strong></dt>
218 <dd>
219 <p>the largest of the data points is stored.</p>
220 </dd>
221 <dt><strong><a name="last" class="item">LAST</a></strong></dt>
223 <dd>
224 <p>the last data points is used.</p>
225 </dd>
226 </dl>
227 <p>Note that data aggregation inevitably leads to loss of precision and
228 information. The trick is to pick the aggregate function such that the
229 <em>interesting</em> properties of your data is kept across the aggregation
230 process.</p>
231 <p>The format of <strong>RRA</strong> line for these
232 consolidation functions is:</p>
233 <p><strong>RRA:</strong><em>AVERAGE | MIN | MAX | LAST</em><strong>:</strong><em>xff</em><strong>:</strong><em>steps</em><strong>:</strong><em>rows</em></p>
234 <p><em>xff</em> The xfiles factor defines what part of a consolidation interval may
235 be made up from <em>*UNKNOWN*</em> data while the consolidated value is still
236 regarded as known. It is given as the ratio of allowed <em>*UNKNOWN*</em> PDPs
237 to the number of PDPs in the interval. Thus, it ranges from 0 to 1 (exclusive).</p>
238 <p><em>steps</em> defines how many of these <em>primary data points</em> are used to build
239 a <em>consolidated data point</em> which then goes into the archive.</p>
240 <p><em>rows</em> defines how many generations of data values are kept in an <strong>RRA</strong>.
241 Obviously, this has to be greater than zero.</p>
242 <p>
243 </p>
244 <hr />
245 <h1><a name="aberrant_behavior_detection_with_holt_winters_forecasting">Aberrant Behavior Detection with Holt-Winters Forecasting</a></h1>
246 <p>In addition to the aggregate functions, there are a set of specialized
247 functions that enable <strong>RRDtool</strong> to provide data smoothing (via the
248 Holt-Winters forecasting algorithm), confidence bands, and the
249 flagging aberrant behavior in the data source time series:</p>
250 <ul>
251 <li>
252 <p><strong>RRA:</strong><em>HWPREDICT</em><strong>:</strong><em>rows</em><strong>:</strong><em>alpha</em><strong>:</strong><em>beta</em><strong>:</strong><em>seasonal period</em>[<strong>:</strong><em>rra-num</em>]</p>
253 </li>
254 <li>
255 <p><strong>RRA:</strong><em>MHWPREDICT</em><strong>:</strong><em>rows</em><strong>:</strong><em>alpha</em><strong>:</strong><em>beta</em><strong>:</strong><em>seasonal period</em>[<strong>:</strong><em>rra-num</em>]</p>
256 </li>
257 <li>
258 <p><strong>RRA:</strong><em>SEASONAL</em><strong>:</strong><em>seasonal period</em><strong>:</strong><em>gamma</em><strong>:</strong><em>rra-num</em>[<strong>:smoothing-window=</strong><em>fraction</em>]</p>
259 </li>
260 <li>
261 <p><strong>RRA:</strong><em>DEVSEASONAL</em><strong>:</strong><em>seasonal period</em><strong>:</strong><em>gamma</em><strong>:</strong><em>rra-num</em>[<strong>:smoothing-window=</strong><em>fraction</em>]</p>
262 </li>
263 <li>
264 <p><strong>RRA:</strong><em>DEVPREDICT</em><strong>:</strong><em>rows</em><strong>:</strong><em>rra-num</em></p>
265 </li>
266 <li>
267 <p><strong>RRA:</strong><em>FAILURES</em><strong>:</strong><em>rows</em><strong>:</strong><em>threshold</em><strong>:</strong><em>window length</em><strong>:</strong><em>rra-num</em></p>
268 </li>
269 </ul>
270 <p>These <strong>RRAs</strong> differ from the true consolidation functions in several ways.
271 First, each of the <strong>RRA</strong>s is updated once for every primary data point.
272 Second, these <strong>RRAs</strong> are interdependent. To generate real-time confidence
273 bounds, a matched set of SEASONAL, DEVSEASONAL, DEVPREDICT, and either
274 HWPREDICT or MHWPREDICT must exist. Generating smoothed values of the primary
275 data points requires a SEASONAL <strong>RRA</strong> and either an HWPREDICT or MHWPREDICT 
276 <strong>RRA</strong>. Aberrant behavior detection requires FAILURES, DEVSEASONAL, SEASONAL,
277 and either HWPREDICT or MHWPREDICT.</p>
278 <p>The predicted, or smoothed, values are stored in the HWPREDICT or MHWPREDICT
279 <strong>RRA</strong>. HWPREDICT and MHWPREDICT are actually two variations on the
280 Holt-Winters method. They are interchangeable. Both attempt to decompose data
281 into three components: a baseline, a trend, and a seasonal coefficient.
282 HWPREDICT adds its seasonal coefficient to the baseline to form a prediction, whereas
283 MHWPREDICT multiplies its seasonal coefficient by the baseline to form a
284 prediction. The difference is noticeable when the baseline changes
285 significantly in the course of a season; HWPREDICT will predict the seasonality
286 to stay constant as the baseline changes, but MHWPREDICT will predict the
287 seasonality to grow or shrink in proportion to the baseline. The proper choice
288 of method depends on the thing being modeled. For simplicity, the rest of this
289 discussion will refer to HWPREDICT, but MHWPREDICT may be substituted in its
290 place.</p>
291 <p>The predicted deviations are stored in DEVPREDICT (think a standard deviation
292 which can be scaled to yield a confidence band). The FAILURES <strong>RRA</strong> stores 
293 binary indicators. A 1 marks the indexed observation as failure; that is, the 
294 number of confidence bounds violations in the preceding window of observations 
295 met or exceeded a specified threshold. An example of using these <strong>RRAs</strong> to graph 
296 confidence bounds and failures appears in <a href="././rrdgraph.html">the rrdgraph manpage</a>.</p>
297 <p>The SEASONAL and DEVSEASONAL <strong>RRAs</strong> store the seasonal coefficients for the
298 Holt-Winters forecasting algorithm and the seasonal deviations, respectively.
299 There is one entry per observation time point in the seasonal cycle. For
300 example, if primary data points are generated every five minutes and the
301 seasonal cycle is 1 day, both SEASONAL and DEVSEASONAL will have 288 rows.</p>
302 <p>In order to simplify the creation for the novice user, in addition to
303 supporting explicit creation of the HWPREDICT, SEASONAL, DEVPREDICT,
304 DEVSEASONAL, and FAILURES <strong>RRAs</strong>, the <strong>RRDtool</strong> create command supports
305 implicit creation of the other four when HWPREDICT is specified alone and
306 the final argument <em>rra-num</em> is omitted.</p>
307 <p><em>rows</em> specifies the length of the <strong>RRA</strong> prior to wrap around. Remember
308 that there is a one-to-one correspondence between primary data points and
309 entries in these RRAs. For the HWPREDICT CF, <em>rows</em> should be larger than
310 the <em>seasonal period</em>. If the DEVPREDICT <strong>RRA</strong> is implicitly created, the
311 default number of rows is the same as the HWPREDICT <em>rows</em> argument. If the
312 FAILURES <strong>RRA</strong> is implicitly created, <em>rows</em> will be set to the <em>seasonal
313 period</em> argument of the HWPREDICT <strong>RRA</strong>. Of course, the <strong>RRDtool</strong>
314 <em>resize</em> command is available if these defaults are not sufficient and the
315 creator wishes to avoid explicit creations of the other specialized function
316 <strong>RRAs</strong>.</p>
317 <p><em>seasonal period</em> specifies the number of primary data points in a seasonal
318 cycle. If SEASONAL and DEVSEASONAL are implicitly created, this argument for
319 those <strong>RRAs</strong> is set automatically to the value specified by HWPREDICT. If
320 they are explicitly created, the creator should verify that all three
321 <em>seasonal period</em> arguments agree.</p>
322 <p><em>alpha</em> is the adaption parameter of the intercept (or baseline)
323 coefficient in the Holt-Winters forecasting algorithm. See <a href="././rrdtool.html">the rrdtool manpage</a> for a
324 description of this algorithm. <em>alpha</em> must lie between 0 and 1. A value
325 closer to 1 means that more recent observations carry greater weight in
326 predicting the baseline component of the forecast. A value closer to 0 means
327 that past history carries greater weight in predicting the baseline
328 component.</p>
329 <p><em>beta</em> is the adaption parameter of the slope (or linear trend) coefficient
330 in the Holt-Winters forecasting algorithm. <em>beta</em> must lie between 0 and 1
331 and plays the same role as <em>alpha</em> with respect to the predicted linear
332 trend.</p>
333 <p><em>gamma</em> is the adaption parameter of the seasonal coefficients in the
334 Holt-Winters forecasting algorithm (HWPREDICT) or the adaption parameter in
335 the exponential smoothing update of the seasonal deviations. It must lie
336 between 0 and 1. If the SEASONAL and DEVSEASONAL <strong>RRAs</strong> are created
337 implicitly, they will both have the same value for <em>gamma</em>: the value
338 specified for the HWPREDICT <em>alpha</em> argument. Note that because there is
339 one seasonal coefficient (or deviation) for each time point during the
340 seasonal cycle, the adaptation rate is much slower than the baseline. Each
341 seasonal coefficient is only updated (or adapts) when the observed value
342 occurs at the offset in the seasonal cycle corresponding to that
343 coefficient.</p>
344 <p>If SEASONAL and DEVSEASONAL <strong>RRAs</strong> are created explicitly, <em>gamma</em> need not
345 be the same for both. Note that <em>gamma</em> can also be changed via the
346 <strong>RRDtool</strong> <em>tune</em> command.</p>
347 <p><em>smoothing-window</em> specifies the fraction of a season that should be
348 averaged around each point. By default, the value of <em>smoothing-window</em> is
349 0.05, which means each value in SEASONAL and DEVSEASONAL will be occasionally
350 replaced by averaging it with its (<em>seasonal period</em>*0.05) nearest neighbors.
351 Setting <em>smoothing-window</em> to zero will disable the running-average smoother
352 altogether.</p>
353 <p><em>rra-num</em> provides the links between related <strong>RRAs</strong>. If HWPREDICT is
354 specified alone and the other <strong>RRAs</strong> are created implicitly, then
355 there is no need to worry about this argument. If <strong>RRAs</strong> are created
356 explicitly, then carefully pay attention to this argument. For each
357 <strong>RRA</strong> which includes this argument, there is a dependency between
358 that <strong>RRA</strong> and another <strong>RRA</strong>. The <em>rra-num</em> argument is the 1-based
359 index in the order of <strong>RRA</strong> creation (that is, the order they appear
360 in the <em>create</em> command). The dependent <strong>RRA</strong> for each <strong>RRA</strong>
361 requiring the <em>rra-num</em> argument is listed here:</p>
362 <ul>
363 <li>
364 <p>HWPREDICT <em>rra-num</em> is the index of the SEASONAL <strong>RRA</strong>.</p>
365 </li>
366 <li>
367 <p>SEASONAL <em>rra-num</em> is the index of the HWPREDICT <strong>RRA</strong>.</p>
368 </li>
369 <li>
370 <p>DEVPREDICT <em>rra-num</em> is the index of the DEVSEASONAL <strong>RRA</strong>.</p>
371 </li>
372 <li>
373 <p>DEVSEASONAL <em>rra-num</em> is the index of the HWPREDICT <strong>RRA</strong>.</p>
374 </li>
375 <li>
376 <p>FAILURES <em>rra-num</em> is the index of the DEVSEASONAL <strong>RRA</strong>.</p>
377 </li>
378 </ul>
379 <p><em>threshold</em> is the minimum number of violations (observed values outside
380 the confidence bounds) within a window that constitutes a failure. If the
381 FAILURES <strong>RRA</strong> is implicitly created, the default value is 7.</p>
382 <p><em>window length</em> is the number of time points in the window. Specify an
383 integer greater than or equal to the threshold and less than or equal to 28.
384 The time interval this window represents depends on the interval between
385 primary data points. If the FAILURES <strong>RRA</strong> is implicitly created, the
386 default value is 9.</p>
387 <p>
388 </p>
389 <hr />
390 <h1><a name="the_heartbeat_and_the_step">The HEARTBEAT and the STEP</a></h1>
391 <p>Here is an explanation by Don Baarda on the inner workings of RRDtool.
392 It may help you to sort out why all this *UNKNOWN* data is popping
393 up in your databases:</p>
394 <p>RRDtool gets fed samples/updates at arbitrary times. From these it builds Primary
395 Data Points (PDPs) on every &quot;step&quot; interval. The PDPs are
396 then accumulated into the RRAs.</p>
397 <p>The &quot;heartbeat&quot; defines the maximum acceptable interval between
398 samples/updates. If the interval between samples is less than &quot;heartbeat&quot;,
399 then an average rate is calculated and applied for that interval. If
400 the interval between samples is longer than &quot;heartbeat&quot;, then that
401 entire interval is considered &quot;unknown&quot;. Note that there are other
402 things that can make a sample interval &quot;unknown&quot;, such as the rate
403 exceeding limits, or a sample that was explicitly marked as unknown.</p>
404 <p>The known rates during a PDP's &quot;step&quot; interval are used to calculate
405 an average rate for that PDP. If the total &quot;unknown&quot; time accounts for
406 more than <strong>half</strong> the &quot;step&quot;, the entire PDP is marked
407 as &quot;unknown&quot;. This means that a mixture of known and &quot;unknown&quot; sample
408 times in a single PDP &quot;step&quot; may or may not add up to enough &quot;known&quot;
409 time to warrent for a known PDP.</p>
410 <p>The &quot;heartbeat&quot; can be short (unusual) or long (typical) relative to
411 the &quot;step&quot; interval between PDPs. A short &quot;heartbeat&quot; means you
412 require multiple samples per PDP, and if you don't get them mark the
413 PDP unknown. A long heartbeat can span multiple &quot;steps&quot;, which means
414 it is acceptable to have multiple PDPs calculated from a single
415 sample. An extreme example of this might be a &quot;step&quot; of 5 minutes and a
416 &quot;heartbeat&quot; of one day, in which case a single sample every day will
417 result in all the PDPs for that entire day period being set to the
418 same average rate. <em>-- Don Baarda &lt;<a href="mailto:don.baarda@baesystems.com">don.baarda@baesystems.com</a>&gt;</em></p>
419 <pre>
420        time|
421        axis|
422  begin__|00|
423         |01|
424        u|02|----* sample1, restart &quot;hb&quot;-timer
425        u|03|   /
426        u|04|  /
427        u|05| /
428        u|06|/     &quot;hbt&quot; expired
429        u|07|
430         |08|----* sample2, restart &quot;hb&quot; 
431         |09|   / 
432         |10|  /
433        u|11|----* sample3, restart &quot;hb&quot;
434        u|12|   /
435        u|13|  /
436  step1_u|14| /
437        u|15|/     &quot;swt&quot; expired
438        u|16|
439         |17|----* sample4, restart &quot;hb&quot;, create &quot;pdp&quot; for step1 = 
440         |18|   /  = unknown due to 10 &quot;u&quot; labled secs &gt; 0.5 * step
441         |19|  /
442         |20| /
443         |21|----* sample5, restart &quot;hb&quot;
444         |22|   /
445         |23|  /
446         |24|----* sample6, restart &quot;hb&quot;
447         |25|   /
448         |26|  /
449         |27|----* sample7, restart &quot;hb&quot;
450  step2__|28|   /
451         |22|  /
452         |23|----* sample8, restart &quot;hb&quot;, create &quot;pdp&quot; for step1, create &quot;cdp&quot; 
453         |24|   /
454         |25|  /</pre>
455 <p>graphics by <em><a href="mailto:vladimir.lavrov@desy.de">vladimir.lavrov@desy.de</a></em>.</p>
456 <p>
457 </p>
458 <hr />
459 <h1><a name="how_to_measure">HOW TO MEASURE</a></h1>
460 <p>Here are a few hints on how to measure:</p>
461 <dl>
462 <dt><strong><a name="temperature" class="item">Temperature</a></strong></dt>
464 <dd>
465 <p>Usually you have some type of meter you can read to get the temperature.
466 The temperature is not really connected with a time. The only connection is
467 that the temperature reading happened at a certain time. You can use the
468 <strong>GAUGE</strong> data source type for this. RRDtool will then record your reading
469 together with the time.</p>
470 </dd>
471 <dt><strong><a name="mail_messages" class="item">Mail Messages</a></strong></dt>
473 <dd>
474 <p>Assume you have a method to count the number of messages transported by
475 your mailserver in a certain amount of time, giving you data like '5
476 messages in the last 65 seconds'. If you look at the count of 5 like an
477 <strong>ABSOLUTE</strong> data type you can simply update the RRD with the number 5 and the
478 end time of your monitoring period. RRDtool will then record the number of
479 messages per second. If at some later stage you want to know the number of
480 messages transported in a day, you can get the average messages per second
481 from RRDtool for the day in question and multiply this number with the
482 number of seconds in a day. Because all math is run with Doubles, the
483 precision should be acceptable.</p>
484 </dd>
485 <dt><strong><a name="it_s_always_a_rate" class="item">It's always a Rate</a></strong></dt>
487 <dd>
488 <p>RRDtool stores rates in amount/second for COUNTER, DERIVE and ABSOLUTE
489 data.  When you plot the data, you will get on the y axis
490 amount/second which you might be tempted to convert to an absolute
491 amount by multiplying by the delta-time between the points. RRDtool
492 plots continuous data, and as such is not appropriate for plotting
493 absolute amounts as for example &quot;total bytes&quot; sent and received in a
494 router. What you probably want is plot rates that you can scale to
495 bytes/hour, for example, or plot absolute amounts with another tool
496 that draws bar-plots, where the delta-time is clear on the plot for
497 each point (such that when you read the graph you see for example GB
498 on the y axis, days on the x axis and one bar for each day).</p>
499 </dd>
500 </dl>
501 <p>
502 </p>
503 <hr />
504 <h1><a name="example">EXAMPLE</a></h1>
505 <pre>
506  rrdtool create temperature.rrd --step 300 \
507   DS:temp:GAUGE:600:-273:5000 \
508   RRA:AVERAGE:0.5:1:1200 \
509   RRA:MIN:0.5:12:2400 \
510   RRA:MAX:0.5:12:2400 \
511   RRA:AVERAGE:0.5:12:2400</pre>
512 <p>This sets up an <strong>RRD</strong> called <em class="file">temperature.rrd</em> which accepts one
513 temperature value every 300 seconds. If no new data is supplied for
514 more than 600 seconds, the temperature becomes <em>*UNKNOWN*</em>.  The
515 minimum acceptable value is -273 and the maximum is 5'000.</p>
516 <p>A few archive areas are also defined. The first stores the
517 temperatures supplied for 100 hours (1'200 * 300 seconds = 100
518 hours). The second RRA stores the minimum temperature recorded over
519 every hour (12 * 300 seconds = 1 hour), for 100 days (2'400 hours). The
520 third and the fourth RRA's do the same for the maximum and
521 average temperature, respectively.</p>
522 <p>
523 </p>
524 <hr />
525 <h1><a name="example_2">EXAMPLE 2</a></h1>
526 <pre>
527  rrdtool create monitor.rrd --step 300        \
528    DS:ifOutOctets:COUNTER:1800:0:4294967295   \
529    RRA:AVERAGE:0.5:1:2016                     \
530    RRA:HWPREDICT:1440:0.1:0.0035:288</pre>
531 <p>This example is a monitor of a router interface. The first <strong>RRA</strong> tracks the
532 traffic flow in octets; the second <strong>RRA</strong> generates the specialized
533 functions <strong>RRAs</strong> for aberrant behavior detection. Note that the <em>rra-num</em>
534 argument of HWPREDICT is missing, so the other <strong>RRAs</strong> will implicitly be
535 created with default parameter values. In this example, the forecasting
536 algorithm baseline adapts quickly; in fact the most recent one hour of
537 observations (each at 5 minute intervals) accounts for 75% of the baseline
538 prediction. The linear trend forecast adapts much more slowly. Observations
539 made during the last day (at 288 observations per day) account for only
540 65% of the predicted linear trend. Note: these computations rely on an
541 exponential smoothing formula described in the LISA 2000 paper.</p>
542 <p>The seasonal cycle is one day (288 data points at 300 second intervals), and
543 the seasonal adaption parameter will be set to 0.1. The RRD file will store 5
544 days (1'440 data points) of forecasts and deviation predictions before wrap
545 around. The file will store 1 day (a seasonal cycle) of 0-1 indicators in
546 the FAILURES <strong>RRA</strong>.</p>
547 <p>The same RRD file and <strong>RRAs</strong> are created with the following command,
548 which explicitly creates all specialized function <strong>RRAs</strong>.</p>
549 <pre>
550  rrdtool create monitor.rrd --step 300 \
551    DS:ifOutOctets:COUNTER:1800:0:4294967295 \
552    RRA:AVERAGE:0.5:1:2016 \
553    RRA:HWPREDICT:1440:0.1:0.0035:288:3 \
554    RRA:SEASONAL:288:0.1:2 \
555    RRA:DEVPREDICT:1440:5 \
556    RRA:DEVSEASONAL:288:0.1:2 \
557    RRA:FAILURES:288:7:9:5</pre>
558 <p>Of course, explicit creation need not replicate implicit create, a
559 number of arguments could be changed.</p>
560 <p>
561 </p>
562 <hr />
563 <h1><a name="example_3">EXAMPLE 3</a></h1>
564 <pre>
565  rrdtool create proxy.rrd --step 300 \
566    DS:Total:DERIVE:1800:0:U  \
567    DS:Duration:DERIVE:1800:0:U  \
568    DS:AvgReqDur:COMPUTE:Duration,Requests,0,EQ,1,Requests,IF,/ \
569    RRA:AVERAGE:0.5:1:2016</pre>
570 <p>This example is monitoring the average request duration during each 300 sec
571 interval for requests processed by a web proxy during the interval.
572 In this case, the proxy exposes two counters, the number of requests
573 processed since boot and the total cumulative duration of all processed
574 requests. Clearly these counters both have some rollover point, but using the
575 DERIVE data source also handles the reset that occurs when the web proxy is
576 stopped and restarted.</p>
577 <p>In the <strong>RRD</strong>, the first data source stores the requests per second rate
578 during the interval. The second data source stores the total duration of all
579 requests processed during the interval divided by 300. The COMPUTE data source
580 divides each PDP of the AccumDuration by the corresponding PDP of
581 TotalRequests and stores the average request duration. The remainder of the
582 RPN expression handles the divide by zero case.</p>
583 <p>
584 </p>
585 <hr />
586 <h1><a name="author">AUTHOR</a></h1>
587 <p>Tobias Oetiker &lt;<a href="mailto:tobi@oetiker.ch">tobi@oetiker.ch</a>&gt;</p>
589 </body>
591 </html>