1 .TH HBAL 1 2009-03-14 htools "Ganeti H-tools"
3 hbal \- Cluster balancer for Ganeti
11 .BI "[ -m " cluster "]"
12 .BI "[-n " nodes-file " ]"
13 .BI "[ -i " instances-file "]"
19 hbal is a cluster balancer that looks at the current state of the
20 cluster (nodes with their total and free disk, memory, etc.) and
21 instance placement and computes a series of steps designed to bring
22 the cluster into a better state.
24 The algorithm to do so is designed to be stable (i.e. it will give you
25 the same results when restarting it from the middle of the solution)
26 and reasonably fast. It is not, however, designed to be a perfect
27 algorithm - it is possible to make it go into a corner from which it
28 can find no improvement, because it only look one "step" ahead.
30 By default, the program will show the solution incrementally as it is
31 computed, in a somewhat cryptic format; for getting the actual Ganeti
32 command list, use the \fB-C\fR option.
36 The program works in independent steps; at each step, we compute the
37 best instance move that lowers the cluster score.
39 The possible move type for an instance are combinations of
40 failover/migrate and replace-disks such that we change one of the
41 instance nodes, and the other one remains (but possibly with changed
42 role, e.g. from primary it becomes secondary). The list is:
44 - replace secondary (r)
45 - replace primary, a composite move (f, r, f)
46 - failover and replace secondary, also composite (f, r)
47 - replace secondary and failover, also composite (r, f)
49 We don't do the only remaining possibility of replacing both nodes
50 (r,f,r,f or the equivalent f,r,f,r) since these move needs an
51 exhaustive search over both candidate primary and secondary nodes, and
52 is O(n*n) in the number of nodes. Furthermore, it doesn't seems to
53 give better scores but will result in more disk replacements.
57 As said before, the algorithm tries to minimise the cluster score at
58 each step. Currently this score is computed as a sum of the following
60 - coefficient of variance of the percent of free memory
61 - coefficient of variance of the percent of reserved memory
62 - coefficient of variance of the percent of free disk
63 - percentage of nodes failing N+1 check
65 The free memory and free disk values help ensure that all nodes are
66 somewhat balanced in their resource usage. The reserved memory helps
67 to ensure that nodes are somewhat balanced in holding secondary
68 instances, and that no node keeps too much memory reserved for
69 N+1. And finally, the N+1 percentage helps guide the algorithm towards
70 eliminating N+1 failures, if possible.
72 Except for the N+1 failures, we use the coefficient of variance since
73 this brings the values into the same unit so to speak, and with a
74 restrict domain of values (between zero and one). The percentage of
75 N+1 failures, while also in this numeric range, doesn't actually has
76 the same meaning, but it has shown to work well.
78 The other alternative, using for N+1 checks the coefficient of
79 variance of (N+1 fail=1, N+1 pass=0) across nodes could hint the
80 algorithm to make more N+1 failures if most nodes are N+1 fail
81 already. Since this (making N+1 failures) is not allowed by other
82 rules of the algorithm, so the N+1 checks would simply not work
85 On a perfectly balanced cluster (all nodes the same size, all
86 instances the same size and spread across the nodes equally), all
87 values would be zero. This doesn't happen too often in practice :)
89 .SS OTHER POSSIBLE METRICS
91 It would be desirable to add more metrics to the algorithm, especially
92 dynamically-computed metrics, such as:
93 - CPU usage of instances, combined with VCPU versus PCPU count
98 The options that can be passed to the program are as follows:
100 .B -C, --print-commands
101 Print the command list at the end of the run. Without this, the
102 program will only show a shorter, but cryptic output.
105 Prints the before and after node status, in a format designed to allow
106 the user to understand the node's most important parameters.
108 The node list will contain these informations:
109 - a character denoting the status of the node, with '-' meaning an
110 offline node, '*' meaning N+1 failure and blank meaning a good
113 - the total node memory
114 - the memory used by the node itself
115 - the free node memory
116 - the reserved node memory, which is the amount of free memory
117 needed for N+1 compliance
120 - number of primary instances
121 - number of secondary instances
122 - percent of free memory
123 - percent of free disk
127 Only shows a one-line output from the program, designed for the case
128 when one wants to look at multiple clusters at once and check their
131 The line will contain four fields:
132 - initial cluster score
133 - number of steps in the solution
134 - final cluster score
135 - improvement in the cluster score
138 .BI "-n" nodefile ", --nodes=" nodefile
139 The name of the file holding node information (if not collecting via
140 RAPI), instead of the default
145 .BI "-i" instancefile ", --instances=" instancefile
146 The name of the file holding instance information (if not collecting
147 via RAPI), instead of the default
153 Collect data not from files but directly from the
155 given as an argument via RAPI. This work for both Ganeti 1.2 and
159 .BI "-l" N ", --max-length=" N
160 Restrict the solution to this length. This can be used for example to
161 automate the execution of the balancing.
165 Increase the output verbosity. Each usage of this option will increase
166 the verbosity (currently more than 2 doesn't make sense) from the
171 Just show the program version and exit.
175 The exist status of the command will be zero, unless for some reason
176 the algorithm fatally failed (e.g. wrong node or instance data).
180 The program does not check its input data for consistency, and aborts
181 with cryptic errors messages in this case.
183 The algorithm is not perfect.
185 The output format is not easily scriptable, and the program should
186 feed moves directly into Ganeti (either via RAPI or via a gnt-debug
193 With the default options, the program shows each individual step and
194 the improvements it brings in cluster score:
199 Loaded 20 nodes, 80 instances
200 Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
201 Initial score: 0.52329131
202 Trying to minimize the CV...
203 1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
204 2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
205 3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
206 4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
207 5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
208 6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
209 7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
210 8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
211 9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
212 10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
213 11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
214 12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
215 13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
216 14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
217 15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
218 Cluster score improved from 0.52329131 to 0.00252594
222 In the above output, we can see:
223 - the input data (here from files) shows a cluster with 20 nodes and
225 - the cluster is not initially N+1 compliant
226 - the initial score is 0.52329131
228 The step list follows, showing the instance, its initial
229 primary/secondary nodes, the new primary secondary, the cluster list,
230 and the actions taken in this step (with 'f' denoting failover/migrate
231 and 'r' denoting replace secondary).
233 Finally, the program shows the improvement in cluster score.
235 A more detailed output is obtained via the \fB-C\fR and \fB-p\fR options:
240 Loaded 20 nodes, 80 instances
241 Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
242 Initial cluster status:
243 N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
244 * node1 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
245 node2 32762 31280 12000 1861 1026 0 8 0.95476 0.55179
246 * node3 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
247 * node4 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
248 * node5 32762 1280 6000 1861 978 5 5 0.03907 0.52573
249 * node6 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
250 * node7 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
251 node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
252 node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
253 * node10 32762 7280 12000 1861 1026 4 4 0.22221 0.55179
254 node11 32762 7280 6000 1861 922 4 5 0.22221 0.49577
255 node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
256 node13 32762 7280 6000 1861 922 4 5 0.22221 0.49577
257 node14 32762 7280 6000 1861 922 4 5 0.22221 0.49577
258 * node15 32762 7280 12000 1861 1131 4 3 0.22221 0.60782
259 node16 32762 31280 0 1861 1860 0 0 0.95476 1.00000
260 node17 32762 7280 6000 1861 1106 5 3 0.22221 0.59479
261 * node18 32762 1280 6000 1396 561 5 3 0.03907 0.40239
262 * node19 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
263 node20 32762 13280 12000 1861 689 3 9 0.40535 0.37068
265 Initial score: 0.52329131
266 Trying to minimize the CV...
267 1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
268 2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
269 3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
270 4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
271 5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
272 6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
273 7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
274 8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
275 9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
276 10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
277 11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
278 12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
279 13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
280 14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
281 15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
282 Cluster score improved from 0.52329131 to 0.00252594
284 Commands to run to reach the above solution:
286 echo gnt-instance migrate instance14
287 echo gnt-instance replace-disks -n node16 instance14
288 echo gnt-instance migrate instance14
290 echo gnt-instance migrate instance54
291 echo gnt-instance replace-disks -n node16 instance54
292 echo gnt-instance migrate instance54
294 echo gnt-instance migrate instance4
295 echo gnt-instance replace-disks -n node16 instance4
297 echo gnt-instance replace-disks -n node2 instance48
298 echo gnt-instance migrate instance48
300 echo gnt-instance replace-disks -n node16 instance93
301 echo gnt-instance migrate instance93
303 echo gnt-instance replace-disks -n node2 instance89
304 echo gnt-instance migrate instance89
306 echo gnt-instance replace-disks -n node16 instance5
307 echo gnt-instance migrate instance5
309 echo gnt-instance migrate instance94
310 echo gnt-instance replace-disks -n node16 instance94
312 echo gnt-instance migrate instance44
313 echo gnt-instance replace-disks -n node15 instance44
315 echo gnt-instance replace-disks -n node16 instance62
317 echo gnt-instance replace-disks -n node16 instance13
319 echo gnt-instance replace-disks -n node7 instance19
321 echo gnt-instance replace-disks -n node1 instance43
323 echo gnt-instance replace-disks -n node4 instance1
325 echo gnt-instance replace-disks -n node17 instance58
327 Final cluster status:
328 N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
329 node1 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
330 node2 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
331 node3 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
332 node4 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
333 node5 32762 7280 6000 1861 1078 4 5 0.22221 0.57947
334 node6 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
335 node7 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
336 node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
337 node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
338 node10 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
339 node11 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
340 node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
341 node13 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
342 node14 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
343 node15 32762 7280 6000 1861 1031 4 4 0.22221 0.55408
344 node16 32762 7280 6000 1861 1060 4 4 0.22221 0.57007
345 node17 32762 7280 6000 1861 1006 5 4 0.22221 0.54105
346 node18 32762 7280 6000 1396 761 4 2 0.22221 0.54570
347 node19 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
348 node20 32762 13280 6000 1861 1089 3 5 0.40535 0.58565
353 Here we see, beside the step list, the initial and final cluster
354 status, with the final one showing all nodes being N+1 compliant, and
355 the command list to reach the final solution. In the initial listing,
356 we see which nodes are not N+1 compliant.
358 The algorithm is stable as long as each step above is fully completed,
359 e.g. in step 8, both the migrate and the replace-disks are
360 done. Otherwise, if only the migrate is done, the input data is
361 changed in a way that the program will output a different solution
362 list (but hopefully will end in the same state).
365 hn1(1), ganeti(7), gnt-instance(8), gnt-node(8)