hbal - Cluster balancer for Ganeti
hbal {backend options...} [algorithm options...] [reporting options...]
hbal –version
Backend options:
{ -m cluster | -L[ path ] [-X] | -t data-file | -I path }
Algorithm options:
[ –max-cpu *cpu-ratio* ] [ –min-disk *disk-ratio* ] [ -l *limit* ] [ -e *score* ] [ -g *delta* ] [ –min-gain-limit *threshold* ] [ -O *name...* ] [ –no-disk-moves ] [ –avoid-disk-moves *factor* ] [ –no-instance-moves ] [ -U *util-file* ] [ –idle-default ] [ –ignore-dynu ] [ –ignore-soft-errors ] [ –mond *yes|no* ] [ –mond-xen ] [ –mond-kvm-rss ] [ –exit-on-missing-mond-data ] [ –evac-mode ] [ –restricted-migration ] [ –select-instances *inst...* ] [ –exclude-instances *inst...* ]
Reporting options:
[ -C[ *file* ] ] [ -p[ *fields* ] ] [ –print-instances ] [ -S *file* ] [ -v... | -q ]
hbal is a cluster balancer that looks at the current state of the cluster (nodes with their total and free disk, memory, etc.) and instance placement and computes a series of steps designed to bring the cluster into a better state.
The algorithm used is designed to be stable (i.e. it will give you the same results when restarting it from the middle of the solution) and reasonably fast. It is not, however, designed to be a perfect algorithm: it is possible to make it go into a corner from which it can find no improvement, because it looks only one “step” ahead.
The program accesses the cluster state via Rapi or Luxi. It also requests data over the network from all MonDs with the –mond option. Currently it uses only data produced by CPUload collector.
By default, the program will show the solution incrementally as it is computed, in a somewhat cryptic format; for getting the actual Ganeti command list, use the -C option.
The program works in independent steps; at each step, we compute the best instance move that lowers the cluster score.
The possible move type for an instance are combinations of failover/migrate and replace-disks such that we change one of the instance nodes, and the other one remains (but possibly with changed role, e.g. from primary it becomes secondary). The list is:
We don’t do the only remaining possibility of replacing both nodes (r,f,r,f or the equivalent f,r,f,r) since these move needs an exhaustive search over both candidate primary and secondary nodes, and is O(n*n) in the number of nodes. Furthermore, it doesn’t seems to give better scores but will result in more disk replacements.
At each step, we prevent an instance move if it would cause:
As said before, the algorithm tries to minimise the cluster score at each step. Currently this score is computed as a weighted sum of the following components:
The free memory and free disk values help ensure that all nodes are somewhat balanced in their resource usage. The reserved memory helps to ensure that nodes are somewhat balanced in holding secondary instances, and that no node keeps too much memory reserved for N+1. And finally, the N+1 percentage helps guide the algorithm towards eliminating N+1 failures, if possible.
Except for the N+1 failures, offline instances counts, failure domain violation counts and desired locations count, we use the standard deviation since when used with values within a fixed range (we use percents expressed as values between zero and one) it gives consistent results across all metrics (there are some small issues related to different means, but it works generally well). The ‘count’ type values will have higher score and thus will matter more for balancing; thus these are better for hard constraints (like evacuating nodes and fixing N+1 failures). For example, the offline instances count (i.e. the number of instances living on offline nodes) will cause the algorithm to actively move instances away from offline nodes. This, coupled with the restriction on placement given by offline nodes, will cause evacuation of such nodes.
The dynamic load values need to be read from an external file (Ganeti doesn’t supply them), and are computed for each node as: sum of primary instance cpu load, sum of primary instance memory load, sum of primary and secondary instance disk load (as DRBD generates write load on secondary nodes too in normal case and in degraded scenarios also read load), and sum of primary instance network load. An example of how to generate these values for input to hbal would be to track xm list for instances over a day and by computing the delta of the cpu values, and feed that via the -U option for all instances (and keep the other metrics as one). For the algorithm to work, all that is needed is that the values are consistent for a metric across all instances (e.g. all instances use cpu% to report cpu usage, and not something related to number of CPU seconds used if the CPUs are different), and that they are normalised to between zero and one. Note that it’s recommended to not have zero as the load value for any instance metric since then secondary instances are not well balanced.
The CPUload from MonD’s data collector will be used only if all MonDs are running, otherwise it won’t affect the cluster score. Since we can’t find the CPU load of each instance, we can assume that the CPU load of an instance is proportional to the number of its vcpus. With this heuristic, instances from nodes with high CPU load will tend to move to nodes with less CPU load.
On a perfectly balanced cluster (all nodes the same size, all instances the same size and spread across the nodes equally, all desired locations satisfied), the values for all metrics would be zero, with the exception of the total percentage of reserved memory. This doesn’t happen too often in practice :)
Since current Ganeti versions do not report the memory used by offline (down) instances, ignoring the run status of instances will cause wrong calculations. For this reason, the algorithm subtracts the memory size of down instances from the free node memory of their primary node, in effect simulating the startup of such instances.
Sometimes, administrators want specific instances located in a particular, typically geographic, location. To suppoer this desired location tags are introduced.
If the cluster is tagged htools:desiredlocation:x then tags starting with x are desired location tags. Instances can be assigned tags of the form x that means that instance wants to be placed on a node tagged with a location tag x. (That means that cluster should be tagged htools:nlocation:x too).
Instance pinning is just heuristics, not a hard enforced requirement; it will only be achieved by the cluster metrics favouring such placements.
The exclusion tags mechanism is designed to prevent instances which run the same workload (e.g. two DNS servers) to land on the same node, which would make the respective node a SPOF for the given service.
It works by tagging instances with certain tags and then building exclusion maps based on these. Which tags are actually used is configured either via the command line (option –exclusion-tags) or via adding them to the cluster tags:
Both the above forms mean that two instances both having (e.g.) the tag a:foo or b:bar won’t end on the same node.
If Ganeti is deployed on a heterogeneous cluster, migration might not be possible between all nodes of a node group. One example of such a situation is upgrading the hypervisor node by node. To make hbal aware of those restrictions, the following cluster tags are used.
So in the simple case of a hypervisor upgrade, tagging all the nodes that have been upgraded with a migration tag suffices. In more complicated situations, it is always possible to use a different migration tag for each hypervisor used and explictly state the allowed migration directions by means of htools:allowmigration: tags.
Within a node group, certain nodes might be more likely to fail simultaneously due to a common cause of error (e.g., if they share the same power supply unit). Ganeti can be made aware of thos common causes of failure by means of tags.
Instances with primary and secondary node having a common cause of failure and instances sharing the same exclusion tag with primary nodes having a common failure are considered badly placed. While such placements are always allowed, they count heavily towards the cluster score.
The options that can be passed to the program are as follows:
Print the command list at the end of the run. Without this, the program will only show a shorter, but cryptic output.
Note that the moves list will be split into independent steps, called “jobsets”, but only for visual inspection, not for actually parallelisation. It is not possible to parallelise these directly when executed via “gnt-instance” commands, since a compound command (e.g. failover and replace-disks) must be executed serially. Parallel execution is only possible when using the Luxi backend and the -L option.
The algorithm for splitting the moves into jobsets is by accumulating moves until the next move is touching nodes already touched by the current moves; this means we can’t execute in parallel (due to resource allocation in Ganeti) and thus we start a new jobset.
This option (which can be given multiple times) will mark nodes as being offline. This means a couple of things:
Note that algorithm will also mark as offline any nodes which are reported by RAPI as such, or that have ”?” in file-based input in any numeric fields.
This parameter denotes how much above the N+1 bound the cluster score can for us to be happy with and alters the computation in two ways:
The default value of the parameter is currently 1e-9 (chosen empirically).
This parameter specifies a file holding instance dynamic utilisation information that will be used to tweak the balancing algorithm to equalise load on the nodes (as opposed to static resource usage). The file is in the format “instance_name cpu_util mem_util disk_util net_util” where the “_util” parameters are interpreted as numbers and the instance name must match exactly the instance as read from Ganeti. In case of unknown instance names, the program will abort.
If not given, the default values are one for all metrics and thus dynamic utilisation has only one effect on the algorithm: the equalisation of the secondary instances across nodes (this is the only metric that is not tracked by another, dedicated value, and thus the disk load of instances will cause secondary instance equalisation). Note that value of one will also influence slightly the primary instance count, but that is already tracked via other metrics and thus the influence of the dynamic utilisation will be practically insignificant.
| -X | When using the Luxi backend, hbal can also execute the given commands. The execution method is to execute the individual jobsets (see the -C option for details) in separate stages, aborting if at any time a jobset doesn’t have all jobs successful. Each step in the balancing solution will be translated into exactly one Ganeti job (having between one and three OpCodes), and all the steps in a jobset will be executed in parallel. The jobsets themselves are executed serially. The execution of the job series can be interrupted, see below for signal handling. |
When executing jobs via LUXI (using the -X option), normally hbal will execute all jobs until either one errors out or all the jobs finish successfully.
Since balancing can take a long time, it is possible to stop hbal early in two ways:
Note that in any situation, it’s perfectly safe to kill hbal, either via the above signals or via any other signal (e.g. SIGQUIT, SIGKILL), since the jobs themselves are processed by Ganeti whereas hbal (after submission) only watches their progression. In this case, the user will have to query Ganeti for job results.
The exit status of the command will be zero, unless for some reason the algorithm failed (e.g. wrong node or instance data), invalid command line options, or (in case of job execution) one of the jobs has failed.
Once job execution via Luxi has started (-X), if the balancing was interrupted early (via SIGINT, or via --max-length) but all jobs executed successfully, then the exit status is zero; a non-zero exit code means that the cluster state should be investigated, since a job failed or we couldn’t compute its status and this can also point to a problem on the Ganeti side.
The program does not check all its input data for consistency, and sometime aborts with cryptic errors messages with invalid data.
The algorithm is not perfect.
Note that these examples are not for the latest version (they don’t have full node data).
With the default options, the program shows each individual step and the improvements it brings in cluster score:
$ hbal
Loaded 20 nodes, 80 instances
Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
Initial score: 0.52329131
Trying to minimize the CV...
1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
Cluster score improved from 0.52329131 to 0.00252594
In the above output, we can see:
The step list follows, showing the instance, its initial primary/secondary nodes, the new primary secondary, the cluster list, and the actions taken in this step (with ‘f’ denoting failover/migrate and ‘r’ denoting replace secondary).
Finally, the program shows the improvement in cluster score.
A more detailed output is obtained via the -C and -p options:
$ hbal
Loaded 20 nodes, 80 instances
Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
Initial cluster status:
N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
* node1 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
node2 32762 31280 12000 1861 1026 0 8 0.95476 0.55179
* node3 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
* node4 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
* node5 32762 1280 6000 1861 978 5 5 0.03907 0.52573
* node6 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
* node7 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
* node10 32762 7280 12000 1861 1026 4 4 0.22221 0.55179
node11 32762 7280 6000 1861 922 4 5 0.22221 0.49577
node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node13 32762 7280 6000 1861 922 4 5 0.22221 0.49577
node14 32762 7280 6000 1861 922 4 5 0.22221 0.49577
* node15 32762 7280 12000 1861 1131 4 3 0.22221 0.60782
node16 32762 31280 0 1861 1860 0 0 0.95476 1.00000
node17 32762 7280 6000 1861 1106 5 3 0.22221 0.59479
* node18 32762 1280 6000 1396 561 5 3 0.03907 0.40239
* node19 32762 1280 6000 1861 1026 5 3 0.03907 0.55179
node20 32762 13280 12000 1861 689 3 9 0.40535 0.37068
Initial score: 0.52329131
Trying to minimize the CV...
1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
Cluster score improved from 0.52329131 to 0.00252594
Commands to run to reach the above solution:
echo step 1
echo gnt-instance migrate instance14
echo gnt-instance replace-disks -n node16 instance14
echo gnt-instance migrate instance14
echo step 2
echo gnt-instance migrate instance54
echo gnt-instance replace-disks -n node16 instance54
echo gnt-instance migrate instance54
echo step 3
echo gnt-instance migrate instance4
echo gnt-instance replace-disks -n node16 instance4
echo step 4
echo gnt-instance replace-disks -n node2 instance48
echo gnt-instance migrate instance48
echo step 5
echo gnt-instance replace-disks -n node16 instance93
echo gnt-instance migrate instance93
echo step 6
echo gnt-instance replace-disks -n node2 instance89
echo gnt-instance migrate instance89
echo step 7
echo gnt-instance replace-disks -n node16 instance5
echo gnt-instance migrate instance5
echo step 8
echo gnt-instance migrate instance94
echo gnt-instance replace-disks -n node16 instance94
echo step 9
echo gnt-instance migrate instance44
echo gnt-instance replace-disks -n node15 instance44
echo step 10
echo gnt-instance replace-disks -n node16 instance62
echo step 11
echo gnt-instance replace-disks -n node16 instance13
echo step 12
echo gnt-instance replace-disks -n node7 instance19
echo step 13
echo gnt-instance replace-disks -n node1 instance43
echo step 14
echo gnt-instance replace-disks -n node4 instance1
echo step 15
echo gnt-instance replace-disks -n node17 instance58
Final cluster status:
N1 Name t_mem f_mem r_mem t_dsk f_dsk pri sec p_fmem p_fdsk
node1 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node2 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node3 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node4 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node5 32762 7280 6000 1861 1078 4 5 0.22221 0.57947
node6 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node7 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node8 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node9 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node10 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node11 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
node12 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node13 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
node14 32762 7280 6000 1861 1022 4 4 0.22221 0.54951
node15 32762 7280 6000 1861 1031 4 4 0.22221 0.55408
node16 32762 7280 6000 1861 1060 4 4 0.22221 0.57007
node17 32762 7280 6000 1861 1006 5 4 0.22221 0.54105
node18 32762 7280 6000 1396 761 4 2 0.22221 0.54570
node19 32762 7280 6000 1861 1026 4 4 0.22221 0.55179
node20 32762 13280 6000 1861 1089 3 5 0.40535 0.58565
Here we see, beside the step list, the initial and final cluster status, with the final one showing all nodes being N+1 compliant, and the command list to reach the final solution. In the initial listing, we see which nodes are not N+1 compliant.
The algorithm is stable as long as each step above is fully completed, e.g. in step 8, both the migrate and the replace-disks are done. Otherwise, if only the migrate is done, the input data is changed in a way that the program will output a different solution list (but hopefully will end in the same state).