1 .TH HBAL 1 2009-03-23 htools "Ganeti H-tools"
3 hbal \- Cluster balancer for Ganeti
12 .BI "[-m " cluster "]"
13 .BI "[-n " nodes-file " ]"
14 .BI "[-i " instances-file "]"
20 hbal is a cluster balancer that looks at the current state of the
21 cluster (nodes with their total and free disk, memory, etc.) and
22 instance placement and computes a series of steps designed to bring
23 the cluster into a better state.
25 The algorithm to do so is designed to be stable (i.e. it will give you
26 the same results when restarting it from the middle of the solution)
27 and reasonably fast. It is not, however, designed to be a perfect
28 algorithm - it is possible to make it go into a corner from which it
29 can find no improvement, because it only look one "step" ahead.
31 By default, the program will show the solution incrementally as it is
32 computed, in a somewhat cryptic format; for getting the actual Ganeti
33 command list, use the \fB-C\fR option.
37 The program works in independent steps; at each step, we compute the
38 best instance move that lowers the cluster score.
40 The possible move type for an instance are combinations of
41 failover/migrate and replace-disks such that we change one of the
42 instance nodes, and the other one remains (but possibly with changed
43 role, e.g. from primary it becomes secondary). The list is:
53 replace primary, a composite move (f, r, f)
56 failover and replace secondary, also composite (f, r)
59 replace secondary and failover, also composite (r, f)
62 We don't do the only remaining possibility of replacing both nodes
63 (r,f,r,f or the equivalent f,r,f,r) since these move needs an
64 exhaustive search over both candidate primary and secondary nodes, and
65 is O(n*n) in the number of nodes. Furthermore, it doesn't seems to
66 give better scores but will result in more disk replacements.
70 As said before, the algorithm tries to minimise the cluster score at
71 each step. Currently this score is computed as a sum of the following
76 coefficient of variance of the percent of free memory
79 coefficient of variance of the percent of reserved memory
82 coefficient of variance of the percent of free disk
85 percentage of nodes failing N+1 check
88 percentage of instances living (either as primary or secondary) on
92 The free memory and free disk values help ensure that all nodes are
93 somewhat balanced in their resource usage. The reserved memory helps
94 to ensure that nodes are somewhat balanced in holding secondary
95 instances, and that no node keeps too much memory reserved for
96 N+1. And finally, the N+1 percentage helps guide the algorithm towards
97 eliminating N+1 failures, if possible.
99 Except for the N+1 failures and offline instances percentage, we use
100 the coefficient of variance since this brings the values into the same
101 unit so to speak, and with a restrict domain of values (between zero
102 and one). The percentage of N+1 failures, while also in this numeric
103 range, doesn't actually has the same meaning, but it has shown to work
106 The other alternative, using for N+1 checks the coefficient of
107 variance of (N+1 fail=1, N+1 pass=0) across nodes could hint the
108 algorithm to make more N+1 failures if most nodes are N+1 fail
109 already. Since this (making N+1 failures) is not allowed by other
110 rules of the algorithm, so the N+1 checks would simply not work
111 anymore in this case.
113 The offline instances percentage (meaning the percentage of instances
114 living on offline nodes) will cause the algorithm to actively move
115 instances away from offline nodes. This, coupled with the restriction
116 on placement given by offline nodes, will cause evacuation of such
119 On a perfectly balanced cluster (all nodes the same size, all
120 instances the same size and spread across the nodes equally), all
121 values would be zero. This doesn't happen too often in practice :)
123 .SS OFFLINE INSTANCES
125 Since current Ganeti versions do not report the memory used by offline
126 (down) instances, ignoring the run status of instances will cause
127 wrong calculations. For this reason, the algorithm subtracts the
128 memory size of down instances from the free node memory of their
129 primary node, in effect simulating the startup of such instances.
131 .SS OTHER POSSIBLE METRICS
133 It would be desirable to add more metrics to the algorithm, especially
134 dynamically-computed metrics, such as:
138 CPU usage of instances, combined with VCPU versus PCPU count
148 The options that can be passed to the program are as follows:
150 .B -C, --print-commands
151 Print the command list at the end of the run. Without this, the
152 program will only show a shorter, but cryptic output.
155 Prints the before and after node status, in a format designed to allow
156 the user to understand the node's most important parameters.
158 The node list will contain these informations:
162 a character denoting the status of the node, with '-' meaning an
163 offline node, '*' meaning N+1 failure and blank meaning a good node
169 the total node memory
172 the memory used by the node itself
175 the memory used by instances
178 amount memory which seems to be in use but cannot be determined why or
179 by which instance; usually this means that the hypervisor has some
180 overhead or that there are other reporting errors
186 the reserved node memory, which is the amount of free memory needed
196 number of primary instances
199 number of secondary instances
202 percent of free memory
210 Only shows a one-line output from the program, designed for the case
211 when one wants to look at multiple clusters at once and check their
214 The line will contain four fields:
219 initial cluster score
222 number of steps in the solution
228 improvement in the cluster score
234 This option (which can be given multiple times) will mark nodes as
235 being \fIoffline\fR. This means a couple of things:
240 instances won't be placed on these nodes, not even temporarily;
241 e.g. the \fIreplace primary\fR move is not available if the secondary
242 node is offline, since this move requires a failover.
245 these nodes will not be included in the score calculation (except for
246 the percentage of instances on offline nodes)
251 .BI "-n" nodefile ", --nodes=" nodefile
252 The name of the file holding node information (if not collecting via
253 RAPI), instead of the default
258 .BI "-i" instancefile ", --instances=" instancefile
259 The name of the file holding instance information (if not collecting
260 via RAPI), instead of the default
266 Collect data not from files but directly from the
268 given as an argument via RAPI. This work for both Ganeti 1.2 and
272 .BI "-l" N ", --max-length=" N
273 Restrict the solution to this length. This can be used for example to
274 automate the execution of the balancing.
278 Increase the output verbosity. Each usage of this option will increase
279 the verbosity (currently more than 2 doesn't make sense) from the
284 Just show the program version and exit.
288 The exist status of the command will be zero, unless for some reason
289 the algorithm fatally failed (e.g. wrong node or instance data).
293 The program does not check its input data for consistency, and aborts
294 with cryptic errors messages in this case.
296 The algorithm is not perfect.
298 The algorithm doesn't deal with non-\fBdrbd\fR instances, and chokes
299 on input data which has such instances.
301 The output format is not easily scriptable, and the program should
302 feed moves directly into Ganeti (either via RAPI or via a gnt-debug
307 Note that this example are not for the latest version (they don't have
312 With the default options, the program shows each individual step and
313 the improvements it brings in cluster score:
318 Loaded 20 nodes, 80 instances
319 Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
320 Initial score: 0.52329131
321 Trying to minimize the CV...
322 1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
323 2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
324 3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
325 4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
326 5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
327 6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
328 7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
329 8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
330 9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
331 10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
332 11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
333 12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
334 13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
335 14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
336 15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
337 Cluster score improved from 0.52329131 to 0.00252594
341 In the above output, we can see:
342 - the input data (here from files) shows a cluster with 20 nodes and
344 - the cluster is not initially N+1 compliant
345 - the initial score is 0.52329131
347 The step list follows, showing the instance, its initial
348 primary/secondary nodes, the new primary secondary, the cluster list,
349 and the actions taken in this step (with 'f' denoting failover/migrate
350 and 'r' denoting replace secondary).
352 Finally, the program shows the improvement in cluster score.
354 A more detailed output is obtained via the \fB-C\fR and \fB-p\fR options:
359 Loaded 20 nodes, 80 instances
360 Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
361 Initial cluster status:
362 N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
363 * node1 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
364 node2 32762 31280 12000 1861 1026 0 8 0.95476 0.55179
365 * node3 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
366 * node4 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
367 * node5 32762 1280 6000 1861 978 5 5 0.03907 0.52573
368 * node6 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
369 * node7 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
370 node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
371 node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
372 * node10 32762 7280 12000 1861 1026 4 4 0.22221 0.55179
373 node11 32762 7280 6000 1861 922 4 5 0.22221 0.49577
374 node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
375 node13 32762 7280 6000 1861 922 4 5 0.22221 0.49577
376 node14 32762 7280 6000 1861 922 4 5 0.22221 0.49577
377 * node15 32762 7280 12000 1861 1131 4 3 0.22221 0.60782
378 node16 32762 31280 0 1861 1860 0 0 0.95476 1.00000
379 node17 32762 7280 6000 1861 1106 5 3 0.22221 0.59479
380 * node18 32762 1280 6000 1396 561 5 3 0.03907 0.40239
381 * node19 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
382 node20 32762 13280 12000 1861 689 3 9 0.40535 0.37068
384 Initial score: 0.52329131
385 Trying to minimize the CV...
386 1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
387 2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
388 3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
389 4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
390 5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
391 6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
392 7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
393 8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
394 9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
395 10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
396 11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
397 12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
398 13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
399 14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
400 15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
401 Cluster score improved from 0.52329131 to 0.00252594
403 Commands to run to reach the above solution:
405 echo gnt-instance migrate instance14
406 echo gnt-instance replace-disks -n node16 instance14
407 echo gnt-instance migrate instance14
409 echo gnt-instance migrate instance54
410 echo gnt-instance replace-disks -n node16 instance54
411 echo gnt-instance migrate instance54
413 echo gnt-instance migrate instance4
414 echo gnt-instance replace-disks -n node16 instance4
416 echo gnt-instance replace-disks -n node2 instance48
417 echo gnt-instance migrate instance48
419 echo gnt-instance replace-disks -n node16 instance93
420 echo gnt-instance migrate instance93
422 echo gnt-instance replace-disks -n node2 instance89
423 echo gnt-instance migrate instance89
425 echo gnt-instance replace-disks -n node16 instance5
426 echo gnt-instance migrate instance5
428 echo gnt-instance migrate instance94
429 echo gnt-instance replace-disks -n node16 instance94
431 echo gnt-instance migrate instance44
432 echo gnt-instance replace-disks -n node15 instance44
434 echo gnt-instance replace-disks -n node16 instance62
436 echo gnt-instance replace-disks -n node16 instance13
438 echo gnt-instance replace-disks -n node7 instance19
440 echo gnt-instance replace-disks -n node1 instance43
442 echo gnt-instance replace-disks -n node4 instance1
444 echo gnt-instance replace-disks -n node17 instance58
446 Final cluster status:
447 N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
448 node1 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
449 node2 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
450 node3 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
451 node4 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
452 node5 32762 7280 6000 1861 1078 4 5 0.22221 0.57947
453 node6 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
454 node7 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
455 node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
456 node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
457 node10 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
458 node11 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
459 node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
460 node13 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
461 node14 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
462 node15 32762 7280 6000 1861 1031 4 4 0.22221 0.55408
463 node16 32762 7280 6000 1861 1060 4 4 0.22221 0.57007
464 node17 32762 7280 6000 1861 1006 5 4 0.22221 0.54105
465 node18 32762 7280 6000 1396 761 4 2 0.22221 0.54570
466 node19 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
467 node20 32762 13280 6000 1861 1089 3 5 0.40535 0.58565
472 Here we see, beside the step list, the initial and final cluster
473 status, with the final one showing all nodes being N+1 compliant, and
474 the command list to reach the final solution. In the initial listing,
475 we see which nodes are not N+1 compliant.
477 The algorithm is stable as long as each step above is fully completed,
478 e.g. in step 8, both the migrate and the replace-disks are
479 done. Otherwise, if only the migrate is done, the input data is
480 changed in a way that the program will output a different solution
481 list (but hopefully will end in the same state).
484 .BR hn1 "(1), " hscan "(1), " ganeti "(7), " gnt-instance "(8), "