Synnefo » snf-ganeti

Ganeti Cluster tools (ganeti-htools)

====================================

These are some simple cluster tools for fixing common allocation

problems on Ganeti 2.0 clusters.

Note that these tools are most useful for bigger cluster sizes

(e.g. more than five or ten machines); at lower sizes, the

computations they do can also be done manually.

Most of the tools revolve around the concept of keeping the cluster

N+1 compliant: this means that in case of failure of any node, the

instances affected can be failed over (via ``gnt-node failover`` or

``gnt-instance failover``) to their secondary node, and there is

enough memory reserved for this operation without needing to shutdown

other instances or rebalance the cluster.

**Quick start** (see the installation section for more details):

- (have the ghc compiler and the prerequisite libraries installed)

- make

- ./hbal -m $cluster -C -p

- look at the original and final cluster layout, and if acceptable,

  execute the given commands

Available tools

---------------

Cluster rebalancer

~~~~~~~~~~~~~~~~~~

The rebalancer uses a simple algorithm to try to get the nodes of the

cluster as equal as possible in their resource usage. It tries to

repeatedly move each instance one step, so that the cluster score

becomes better. We stop when no further move can improve the score.

For algorithm details and usage, see the man page hbal(1).

IAllocator plugin

~~~~~~~~~~~~~~~~~

The ``hail`` iallocator plugin can be used for allocations of mirrored

and non-mirrored instances and for relocations of mirrored

instances. It needs to be installed in Ganeti's iallocator search

path—usually ``/usr/lib/ganeti/iallocators`` or

``/usr/local/lib/ganeti/iallocators``, and after that it can be used

via ganeti's ``--iallocator`` option (in various gnt-node/gnt-instance

commands). See the man page hail(1) for more details.

Cluster capacity estimator

~~~~~~~~~~~~~~~~~~~~~~~~~~

The ``hspace`` program will, given an input instance specification,

estimate how many instances of those type can be place on the cluster

before it will become full (as in any new allocation would fail N+1

checks). For more details, see the man page hspace(1).

Integration with Ganeti

-----------------------

The ``hbal`` and ``hspace`` programs can either get their input from

text files, locally from the master daemon (when run on the master

node of a cluster), or remote from a cluster via RAPI. The "-L"

argument enables local collection (with an optional path to the unix

socket). For online collection via RAPI, the "-m" argument should

specify the cluster or master node name. Only ``hbal`` and ``hspace``

use these arguments, ``hail`` uses the standard iallocator API and

thus doesn't need any special setup (just needs to be installed in the

right directory).

For generating the text files, a separate tool (``hscan``) is provided

to automate their gathering if RAPI is available, which is better

since it can extract more precise information. In case RAPI is not

usable for whatever reason, the following two commands should be run::

    gnt-node list -oname,mtotal,mnode,mfree,dtotal,dfree,ctotal,offline \

      --separator '|' --no-headers > nodes

    gnt-instance list -oname,be/memory,sda_size,be/vcpus,status,pnode,snodes \

      --separator '|' --no-head > instances

These two files should be saved under the names of *nodes* and

*instances*.

The ``hail`` program gets its data automatically from Ganeti when used

as described in its section.

Installation

------------

If installing from source, you need a working ghc compiler (6.8 at

least) and some extra Haskell libraries which usually need to be

installed manually:

- json (http://hackage.haskell.org/package/json)

- curl (http://hackage.haskell.org/package/curl)

- network (http://hackage.haskell.org/package/network)

Once these are installed, just typing *make* in the top-level

directory should be enough.

Only the ``hail`` program needs to be installed in a specific place,

the other tools are not location-dependent.

For running the (admittedly small) unittest suite (via *make check*),

the QuickCheck version 1 library is needed.

Internal (implementation) documentation is available in the ``apidoc``

directory.

.. Local Variables:

.. mode: rst

.. End: