1 .TH HBAL 1 2009-03-23 htools "Ganeti H-tools"
3 hbal \- Cluster balancer for Ganeti
14 .BI "[-m " cluster "]"
15 .BI "[-n " nodes-file " ]"
16 .BI "[-i " instances-file "]"
22 hbal is a cluster balancer that looks at the current state of the
23 cluster (nodes with their total and free disk, memory, etc.) and
24 instance placement and computes a series of steps designed to bring
25 the cluster into a better state.
27 The algorithm to do so is designed to be stable (i.e. it will give you
28 the same results when restarting it from the middle of the solution)
29 and reasonably fast. It is not, however, designed to be a perfect
30 algorithm - it is possible to make it go into a corner from which it
31 can find no improvement, because it only look one "step" ahead.
33 By default, the program will show the solution incrementally as it is
34 computed, in a somewhat cryptic format; for getting the actual Ganeti
35 command list, use the \fB-C\fR option.
39 The program works in independent steps; at each step, we compute the
40 best instance move that lowers the cluster score.
42 The possible move type for an instance are combinations of
43 failover/migrate and replace-disks such that we change one of the
44 instance nodes, and the other one remains (but possibly with changed
45 role, e.g. from primary it becomes secondary). The list is:
55 replace primary, a composite move (f, r, f)
58 failover and replace secondary, also composite (f, r)
61 replace secondary and failover, also composite (r, f)
64 We don't do the only remaining possibility of replacing both nodes
65 (r,f,r,f or the equivalent f,r,f,r) since these move needs an
66 exhaustive search over both candidate primary and secondary nodes, and
67 is O(n*n) in the number of nodes. Furthermore, it doesn't seems to
68 give better scores but will result in more disk replacements.
72 As said before, the algorithm tries to minimise the cluster score at
73 each step. Currently this score is computed as a sum of the following
78 coefficient of variance of the percent of free memory
81 coefficient of variance of the percent of reserved memory
84 coefficient of variance of the percent of free disk
87 percentage of nodes failing N+1 check
90 percentage of instances living (either as primary or secondary) on
94 The free memory and free disk values help ensure that all nodes are
95 somewhat balanced in their resource usage. The reserved memory helps
96 to ensure that nodes are somewhat balanced in holding secondary
97 instances, and that no node keeps too much memory reserved for
98 N+1. And finally, the N+1 percentage helps guide the algorithm towards
99 eliminating N+1 failures, if possible.
101 Except for the N+1 failures and offline instances percentage, we use
102 the coefficient of variance since this brings the values into the same
103 unit so to speak, and with a restrict domain of values (between zero
104 and one). The percentage of N+1 failures, while also in this numeric
105 range, doesn't actually has the same meaning, but it has shown to work
108 The other alternative, using for N+1 checks the coefficient of
109 variance of (N+1 fail=1, N+1 pass=0) across nodes could hint the
110 algorithm to make more N+1 failures if most nodes are N+1 fail
111 already. Since this (making N+1 failures) is not allowed by other
112 rules of the algorithm, so the N+1 checks would simply not work
113 anymore in this case.
115 The offline instances percentage (meaning the percentage of instances
116 living on offline nodes) will cause the algorithm to actively move
117 instances away from offline nodes. This, coupled with the restriction
118 on placement given by offline nodes, will cause evacuation of such
121 On a perfectly balanced cluster (all nodes the same size, all
122 instances the same size and spread across the nodes equally), all
123 values would be zero. This doesn't happen too often in practice :)
125 .SS OFFLINE INSTANCES
127 Since current Ganeti versions do not report the memory used by offline
128 (down) instances, ignoring the run status of instances will cause
129 wrong calculations. For this reason, the algorithm subtracts the
130 memory size of down instances from the free node memory of their
131 primary node, in effect simulating the startup of such instances.
133 .SS OTHER POSSIBLE METRICS
135 It would be desirable to add more metrics to the algorithm, especially
136 dynamically-computed metrics, such as:
140 CPU usage of instances, combined with VCPU versus PCPU count
150 The options that can be passed to the program are as follows:
152 .B -C, --print-commands
153 Print the command list at the end of the run. Without this, the
154 program will only show a shorter, but cryptic output.
157 Prints the before and after node status, in a format designed to allow
158 the user to understand the node's most important parameters.
160 The node list will contain these informations:
164 a character denoting the status of the node, with '-' meaning an
165 offline node, '*' meaning N+1 failure and blank meaning a good node
171 the total node memory
174 the memory used by the node itself
177 the memory used by instances
180 amount memory which seems to be in use but cannot be determined why or
181 by which instance; usually this means that the hypervisor has some
182 overhead or that there are other reporting errors
188 the reserved node memory, which is the amount of free memory needed
198 number of primary instances
201 number of secondary instances
204 percent of free memory
212 Only shows a one-line output from the program, designed for the case
213 when one wants to look at multiple clusters at once and check their
216 The line will contain four fields:
221 initial cluster score
224 number of steps in the solution
230 improvement in the cluster score
236 This option (which can be given multiple times) will mark nodes as
237 being \fIoffline\fR. This means a couple of things:
242 instances won't be placed on these nodes, not even temporarily;
243 e.g. the \fIreplace primary\fR move is not available if the secondary
244 node is offline, since this move requires a failover.
247 these nodes will not be included in the score calculation (except for
248 the percentage of instances on offline nodes)
250 Note that hbal will also mark as offline any nodes which are reported
251 by RAPI as such, or that have "?" in file-based input in any numeric
256 .BI "-e" score ", --min-score=" score
257 This parameter denotes the minimum score we are happy with and alters
258 the computation in two ways:
263 if the cluster has the initial score lower than this value, then we
264 don't enter the algorithm at all, and exit with success
267 during the iterative process, if we reach a score lower than this
268 value, we exit the algorithm
270 The default value of the parameter is currently \fI1e-9\fR (chosen
275 .BI "-n" nodefile ", --nodes=" nodefile
276 The name of the file holding node information (if not collecting via
277 RAPI), instead of the default \fInodes\fR file (but see below how to
278 customize the default value via the environment).
281 .BI "-i" instancefile ", --instances=" instancefile
282 The name of the file holding instance information (if not collecting
283 via RAPI), instead of the default \fIinstances\fR file (but see below
284 how to customize the default value via the environment).
288 Collect data not from files but directly from the
290 given as an argument via RAPI. If the argument doesn't contain a colon
291 (:), then it is converted into a fully-built URL via prepending
292 https:// and appending the default RAPI port, otherwise it's
293 considered a fully-specified URL and is used unchanged.
296 .BI "-l" N ", --max-length=" N
297 Restrict the solution to this length. This can be used for example to
298 automate the execution of the balancing.
302 Increase the output verbosity. Each usage of this option will increase
303 the verbosity (currently more than 2 doesn't make sense) from the
308 Decrease the output verbosity. Each usage of this option will decrease
309 the verbosity (less than zero doesn't make sense) from the default of
314 Just show the program version and exit.
318 The exist status of the command will be zero, unless for some reason
319 the algorithm fatally failed (e.g. wrong node or instance data).
323 If the variables \fBHTOOLS_NODES\fR and \fBHTOOLS_INSTANCES\fR are
324 present in the environment, they will override the default names for
325 the nodes and instances files. These will have of course no effect
330 The program does not check its input data for consistency, and aborts
331 with cryptic errors messages in this case.
333 The algorithm is not perfect.
335 The algorithm doesn't deal with non-\fBdrbd\fR instances, and chokes
336 on input data which has such instances.
338 The output format is not easily scriptable, and the program should
339 feed moves directly into Ganeti (either via RAPI or via a gnt-debug
344 Note that this example are not for the latest version (they don't have
349 With the default options, the program shows each individual step and
350 the improvements it brings in cluster score:
355 Loaded 20 nodes, 80 instances
356 Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
357 Initial score: 0.52329131
358 Trying to minimize the CV...
359 1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
360 2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
361 3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
362 4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
363 5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
364 6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
365 7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
366 8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
367 9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
368 10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
369 11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
370 12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
371 13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
372 14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
373 15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
374 Cluster score improved from 0.52329131 to 0.00252594
378 In the above output, we can see:
379 - the input data (here from files) shows a cluster with 20 nodes and
381 - the cluster is not initially N+1 compliant
382 - the initial score is 0.52329131
384 The step list follows, showing the instance, its initial
385 primary/secondary nodes, the new primary secondary, the cluster list,
386 and the actions taken in this step (with 'f' denoting failover/migrate
387 and 'r' denoting replace secondary).
389 Finally, the program shows the improvement in cluster score.
391 A more detailed output is obtained via the \fB-C\fR and \fB-p\fR options:
396 Loaded 20 nodes, 80 instances
397 Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
398 Initial cluster status:
399 N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
400 * node1 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
401 node2 32762 31280 12000 1861 1026 0 8 0.95476 0.55179
402 * node3 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
403 * node4 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
404 * node5 32762 1280 6000 1861 978 5 5 0.03907 0.52573
405 * node6 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
406 * node7 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
407 node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
408 node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
409 * node10 32762 7280 12000 1861 1026 4 4 0.22221 0.55179
410 node11 32762 7280 6000 1861 922 4 5 0.22221 0.49577
411 node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
412 node13 32762 7280 6000 1861 922 4 5 0.22221 0.49577
413 node14 32762 7280 6000 1861 922 4 5 0.22221 0.49577
414 * node15 32762 7280 12000 1861 1131 4 3 0.22221 0.60782
415 node16 32762 31280 0 1861 1860 0 0 0.95476 1.00000
416 node17 32762 7280 6000 1861 1106 5 3 0.22221 0.59479
417 * node18 32762 1280 6000 1396 561 5 3 0.03907 0.40239
418 * node19 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
419 node20 32762 13280 12000 1861 689 3 9 0.40535 0.37068
421 Initial score: 0.52329131
422 Trying to minimize the CV...
423 1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
424 2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
425 3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
426 4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
427 5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
428 6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
429 7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
430 8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
431 9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
432 10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
433 11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
434 12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
435 13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
436 14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
437 15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
438 Cluster score improved from 0.52329131 to 0.00252594
440 Commands to run to reach the above solution:
442 echo gnt-instance migrate instance14
443 echo gnt-instance replace-disks -n node16 instance14
444 echo gnt-instance migrate instance14
446 echo gnt-instance migrate instance54
447 echo gnt-instance replace-disks -n node16 instance54
448 echo gnt-instance migrate instance54
450 echo gnt-instance migrate instance4
451 echo gnt-instance replace-disks -n node16 instance4
453 echo gnt-instance replace-disks -n node2 instance48
454 echo gnt-instance migrate instance48
456 echo gnt-instance replace-disks -n node16 instance93
457 echo gnt-instance migrate instance93
459 echo gnt-instance replace-disks -n node2 instance89
460 echo gnt-instance migrate instance89
462 echo gnt-instance replace-disks -n node16 instance5
463 echo gnt-instance migrate instance5
465 echo gnt-instance migrate instance94
466 echo gnt-instance replace-disks -n node16 instance94
468 echo gnt-instance migrate instance44
469 echo gnt-instance replace-disks -n node15 instance44
471 echo gnt-instance replace-disks -n node16 instance62
473 echo gnt-instance replace-disks -n node16 instance13
475 echo gnt-instance replace-disks -n node7 instance19
477 echo gnt-instance replace-disks -n node1 instance43
479 echo gnt-instance replace-disks -n node4 instance1
481 echo gnt-instance replace-disks -n node17 instance58
483 Final cluster status:
484 N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
485 node1 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
486 node2 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
487 node3 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
488 node4 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
489 node5 32762 7280 6000 1861 1078 4 5 0.22221 0.57947
490 node6 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
491 node7 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
492 node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
493 node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
494 node10 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
495 node11 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
496 node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
497 node13 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
498 node14 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
499 node15 32762 7280 6000 1861 1031 4 4 0.22221 0.55408
500 node16 32762 7280 6000 1861 1060 4 4 0.22221 0.57007
501 node17 32762 7280 6000 1861 1006 5 4 0.22221 0.54105
502 node18 32762 7280 6000 1396 761 4 2 0.22221 0.54570
503 node19 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
504 node20 32762 13280 6000 1861 1089 3 5 0.40535 0.58565
509 Here we see, beside the step list, the initial and final cluster
510 status, with the final one showing all nodes being N+1 compliant, and
511 the command list to reach the final solution. In the initial listing,
512 we see which nodes are not N+1 compliant.
514 The algorithm is stable as long as each step above is fully completed,
515 e.g. in step 8, both the migrate and the replace-disks are
516 done. Otherwise, if only the migrate is done, the input data is
517 changed in a way that the program will output a different solution
518 list (but hopefully will end in the same state).
521 .BR hn1 "(1), " hscan "(1), " ganeti "(7), " gnt-instance "(8), "
526 Copyright (C) 2009 Google Inc. Permission is granted to copy,
527 distribute and/or modify under the terms of the GNU General Public
528 License as published by the Free Software Foundation; either version 2
529 of the License, or (at your option) any later version.
531 On Debian systems, the complete text of the GNU General Public License
532 can be found in /usr/share/common-licenses/GPL.