.TH HBAL 1 2009-03-23 htools "Ganeti H-tools"
.SH NAME
hbal \- Cluster balancer for Ganeti
.SH SYNOPSIS
.B hbal
.B "[backend options...]"
.B "[algorithm options...]"
.B "[reporting options...]"
.B hbal
.B --version
.TP
Backend options:
.BI "[ -m " cluster " ]"
|
.BI "[ -L[" path "] [-X]]"
|
.BI "[ -n " nodes-file " ]"
.BI "[ -i " instances-file " ]"
.TP
Algorithm options:
.BI "[ --max-cpu " cpu-ratio " ]"
.BI "[ --min-disk " disk-ratio " ]"
.BI "[ -l " limit " ]"
.BI "[ -e " score " ]"
.BI "[ -O " name... " ]"
.B "[ --no-disk-moves ]"
.BI "[ -U " util-file " ]"
.TP
Reporting options:
.BI "[ -C[" file "] ]"
.B "[ -p ]"
.B "[ --print-instances ]"
.B "[ -o ]"
.B "[ -v... | -q ]"
.SH DESCRIPTION
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 to do so 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 only look one "step" ahead.
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 \fB-C\fR option.
.SS ALGORITHM
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:
.RS 4
.TP 3
\(em
failover (f)
.TP
\(em
replace secondary (r)
.TP
\(em
replace primary, a composite move (f, r, f)
.TP
\(em
failover and replace secondary, also composite (f, r)
.TP
\(em
replace secondary and failover, also composite (r, f)
.RE
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.
.SS CLUSTER SCORING
As said before, the algorithm tries to minimise the cluster score at
each step. Currently this score is computed as a sum of the following
components:
.RS 4
.TP 3
\(em
coefficient of variance of the percent of free memory
.TP
\(em
coefficient of variance of the percent of reserved memory
.TP
\(em
coefficient of variance of the percent of free disk
.TP
\(em
percentage of nodes failing N+1 check
.TP
\(em
percentage of instances living (either as primary or secondary) on
offline nodes
.TP
\(em
coefficent of variance of the ratio of virtual-to-physical cpus (for
primary instaces of the node)
.TP
\(em
coefficients of variance of the dynamic load on the nodes, for cpus,
memory, disk and network
.RE
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 and offline instances percentage, we use
the coefficient of variance since this brings the values into the same
unit so to speak, and with a restrict domain of values (between zero
and one). The percentage of N+1 failures, while also in this numeric
range, doesn't actually has the same meaning, but it has shown to work
well.
The other alternative, using for N+1 checks the coefficient of
variance of (N+1 fail=1, N+1 pass=0) across nodes could hint the
algorithm to make more N+1 failures if most nodes are N+1 fail
already. Since this (making N+1 failures) is not allowed by other
rules of the algorithm, so the N+1 checks would simply not work
anymore in this case.
The offline instances percentage (meaning the percentage 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 instance over a day and by computing the delta of the cpu
values, and feed that via the \fI-U\fR option for all instances (and
keep the other metrics as zero). 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, but they
could represent network bandwith in Gbps).
On a perfectly balanced cluster (all nodes the same size, all
instances the same size and spread across the nodes equally), all
values would be zero. This doesn't happen too often in practice :)
.SS OFFLINE INSTANCES
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.
.SS OTHER POSSIBLE METRICS
It would be desirable to add more metrics to the algorithm, especially
dynamically-computed metrics, such as:
.RS 4
.TP 3
\(em
CPU usage of instances
.TP
\(em
Disk IO usage
.TP
\(em
Network IO
.RE
.SH OPTIONS
The options that can be passed to the program are as follows:
.TP
.B -C, --print-commands
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
\fI-L\fR 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.
.TP
.B -p, --print-nodes
Prints the before and after node status, in a format designed to allow
the user to understand the node's most important parameters.
.TP
.B --print-instances
Prints the before and after instance map. This is less useful as the
node status, but it can help in understanding instance moves.
The node list will contain these informations:
.RS
.TP
.B F
a character denoting the status of the node, with '-' meaning an
offline node, '*' meaning N+1 failure and blank meaning a good node
.TP
.B Name
the node name
.TP
.B t_mem
the total node memory
.TP
.B n_mem
the memory used by the node itself
.TP
.B i_mem
the memory used by instances
.TP
.B x_mem
amount memory which seems to be in use but cannot be determined why or
by which instance; usually this means that the hypervisor has some
overhead or that there are other reporting errors
.TP
.B f_mem
the free node memory
.TP
.B r_mem
the reserved node memory, which is the amount of free memory needed
for N+1 compliance
.TP
.B t_dsk
total disk
.TP
.B f_dsk
free disk
.TP
.B pcpu
the number of physical cpus on the node
.TP
.B vcpu
the number of virtual cpus allocated to primary instances
.TP
.B pri
number of primary instances
.TP
.B sec
number of secondary instances
.TP
.B p_fmem
percent of free memory
.TP
.B p_fdsk
percent of free disk
.TP
.B r_cpu
ratio of virtual to physical cpus
.TP
.B lCpu
the dynamic CPU load (if the information is available)
.TP
.B lMem
the dynamic memory load (if the information is available)
.TP
.B lDsk
the dynamic disk load (if the information is available)
.TP
.B lNet
the dynamic net load (if the information is available)
.RE
.TP
.B -o, --oneline
Only shows a one-line output from the program, designed for the case
when one wants to look at multiple clusters at once and check their
status.
The line will contain four fields:
.RS
.RS 4
.TP 3
\(em
initial cluster score
.TP
\(em
number of steps in the solution
.TP
\(em
final cluster score
.TP
\(em
improvement in the cluster score
.RE
.RE
.TP
.BI "-O " name
This option (which can be given multiple times) will mark nodes as
being \fIoffline\fR. This means a couple of things:
.RS
.RS 4
.TP 3
\(em
instances won't be placed on these nodes, not even temporarily;
e.g. the \fIreplace primary\fR move is not available if the secondary
node is offline, since this move requires a failover.
.TP
\(em
these nodes will not be included in the score calculation (except for
the percentage of instances on offline nodes)
.RE
Note that hbal 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.
.RE
.TP
.BI "-e" score ", --min-score=" score
This parameter denotes the minimum score we are happy with and alters
the computation in two ways:
.RS
.RS 4
.TP 3
\(em
if the cluster has the initial score lower than this value, then we
don't enter the algorithm at all, and exit with success
.TP
\(em
during the iterative process, if we reach a score lower than this
value, we exit the algorithm
.RE
The default value of the parameter is currently \fI1e-9\fR (chosen
empirically).
.RE
.TP
.BI "--no-disk-moves"
This parameter prevents hbal from using disk move (i.e. "gnt-instance
replace-disks") operations. This will result in a much quicker
balancing, but of course the improvements are limited. It is up to the
user to decide when to use one or another.
.TP
.BI "-U" util-file
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 zero for all metrics and thus
dynamic utilisation has no effect on the balancing algorithm.
.TP
.BI "-n" nodefile ", --nodes=" nodefile
The name of the file holding node information (if not collecting via
RAPI), instead of the default \fInodes\fR file (but see below how to
customize the default value via the environment).
.TP
.BI "-i" instancefile ", --instances=" instancefile
The name of the file holding instance information (if not collecting
via RAPI), instead of the default \fIinstances\fR file (but see below
how to customize the default value via the environment).
.TP
.BI "-m" cluster
Collect data not from files but directly from the
.I cluster
given as an argument via RAPI. If the argument doesn't contain a colon
(:), then it is converted into a fully-built URL via prepending
https:// and appending the default RAPI port, otherwise it's
considered a fully-specified URL and is used as-is.
.TP
.BI "-L[" path "]"
Collect data not from files but directly from the master daemon, which
is to be contacted via the luxi (an internal Ganeti protocol). An
optional \fIpath\fR argument is interpreted as the path to the unix
socket on which the master daemon listens; otherwise, the default path
used by ganeti when installed with "--localstatedir=/var" is used.
.TP
.B "-X"
When using the Luxi backend, hbal can also execute the given
commands. The execution method is to execute the individual jobsets
(see the \fI-C\fR 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.
.TP
.BI "-l" N ", --max-length=" N
Restrict the solution to this length. This can be used for example to
automate the execution of the balancing.
.TP
.BI "--max-cpu " cpu-ratio
The maximum virtual-to-physical cpu ratio, as a floating point number
between zero and one. For example, specifying \fIcpu-ratio\fR as
\fB2.5\fR means that, for a 4-cpu machine, a maximum of 10 virtual
cpus should be allowed to be in use for primary instances. A value of
one doesn't make sense though, as that means no disk space can be used
on it.
.TP
.BI "--min-disk " disk-ratio
The minimum amount of free disk space remaining, as a floating point
number. For example, specifying \fIdisk-ratio\fR as \fB0.25\fR means
that at least one quarter of disk space should be left free on nodes.
.TP
.B -v, --verbose
Increase the output verbosity. Each usage of this option will increase
the verbosity (currently more than 2 doesn't make sense) from the
default of one.
.TP
.B -q, --quiet
Decrease the output verbosity. Each usage of this option will decrease
the verbosity (less than zero doesn't make sense) from the default of
one.
.TP
.B -V, --version
Just show the program version and exit.
.SH EXIT STATUS
The exist status of the command will be zero, unless for some reason
the algorithm fatally failed (e.g. wrong node or instance data).
.SH ENVIRONMENT
If the variables \fBHTOOLS_NODES\fR and \fBHTOOLS_INSTANCES\fR are
present in the environment, they will override the default names for
the nodes and instances files. These will have of course no effect
when RAPI is used.
.SH BUGS
The program does not check its input data for consistency, and aborts
with cryptic errors messages in this case.
The algorithm is not perfect.
The output format is not easily scriptable, and the program should
feed moves directly into Ganeti (either via RAPI or via a gnt-debug
input file).
.SH EXAMPLE
Note that this example are not for the latest version (they don't have
full node data).
.SS Default output
With the default options, the program shows each individual step and
the improvements it brings in cluster score:
.in +4n
.nf
.RB "$" " hbal"
Loaded 20 nodes, 80 instances
Cluster is not N+1 happy, continuing but no guarantee that the cluster will end N+1 happy.
Initial score: 0.52329131
Trying to minimize the CV...
1. instance14 node1:node10 => node16:node10 0.42109120 a=f r:node16 f
2. instance54 node4:node15 => node16:node15 0.31904594 a=f r:node16 f
3. instance4 node5:node2 => node2:node16 0.26611015 a=f r:node16
4. instance48 node18:node20 => node2:node18 0.21361717 a=r:node2 f
5. instance93 node19:node18 => node16:node19 0.16166425 a=r:node16 f
6. instance89 node3:node20 => node2:node3 0.11005629 a=r:node2 f
7. instance5 node6:node2 => node16:node6 0.05841589 a=r:node16 f
8. instance94 node7:node20 => node20:node16 0.00658759 a=f r:node16
9. instance44 node20:node2 => node2:node15 0.00438740 a=f r:node15
10. instance62 node14:node18 => node14:node16 0.00390087 a=r:node16
11. instance13 node11:node14 => node11:node16 0.00361787 a=r:node16
12. instance19 node10:node11 => node10:node7 0.00336636 a=r:node7
13. instance43 node12:node13 => node12:node1 0.00305681 a=r:node1
14. instance1 node1:node2 => node1:node4 0.00263124 a=r:node4
15. instance58 node19:node20 => node19:node17 0.00252594 a=r:node17
Cluster score improved from 0.52329131 to 0.00252594
.fi
.in
In the above output, we can see:
- the input data (here from files) shows a cluster with 20 nodes and
80 instances
- the cluster is not initially N+1 compliant
- the initial score is 0.52329131
The step list follows, showing the instance, its initial
primary/secondary nodes, the new primary secondary, the cluster list,
and the actions taken in this step (with 'f' denoting failover/migrate
and 'r' denoting replace secondary).
Finally, the program shows the improvement in cluster score.
A more detailed output is obtained via the \fB-C\fR and \fB-p\fR options:
.in +4n
.nf
.RB "$" " 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
.fi
.in
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).
.SH SEE ALSO
.BR hspace "(1), " hscan "(1), " hail "(1), "
.BR ganeti "(7), " gnt-instance "(8), " gnt-node "(8)"
.SH "COPYRIGHT"
.PP
Copyright (C) 2009 Google Inc. Permission is granted to copy,
distribute and/or modify under the terms of the GNU General Public
License as published by the Free Software Foundation; either version 2
of the License, or (at your option) any later version.
.PP
On Debian systems, the complete text of the GNU General Public License
can be found in /usr/share/common-licenses/GPL.