[Cluster-devel] [Linux-Cluster] Submitting two new resource plugins to the project

Rafael Micó Miranda rmicmirregs at gmail.com
Mon Jun 8 14:58:34 UTC 2009


Hi,

I have developed a couple of resources for Linux-Cluster (CMAN
+rgmanager) which try to fix some needs I see in Linux-Cluster when
compared with other cluster solution (concretely, Linux-HA a.k.a.
Heartbeat). I am a Linux-HA user and I think this two functionalities
could be useful in Linux-Cluster. 

I would like to give them (both resources) to the community to make them
be into the project, and maybe after testing/quality testing or so be
included into the RedHat Enterprise Linux packages of Linux-Cluster, so
RedHat will give support for them and include them into the
system-config-cluster tool to have a GUI that can configure this
resources and handle their information.

I'll give you some details of both resources:

1.- ping-group: tries to bring to Linux-Cluster the Ping Group
functionality of Linux-HA.

For those who don't know Ping Group, the idea is the following: its a
NODE functionality (not a service or a resource) that checks IP
communications with a list of given client nodes. When failed, Ping
Group will move all services running in the affected node to other nodes
that have proved that keep their communications right, so the service is
provided to the clients even if there is a network problem that affects
only a node of your cluster but the cluster itself wont realize about
it.

I have developed ping-group as a resource to be used into a service of
your cluster, so in the resource arguments you can specify a list of
clients that service should take note on. There is one thing that could
be improved: ping-group will mark the service as failed even if the
other nodes of the cluster would fail too due to lack of communications
with the clients (for example, all clients are powered off). In this
situation the service will go on migrating from one node to another
according to your service failover policy and finally the service will
be stopped. Maybe some ideas could be useful to improve this behaviour.


2.- lvm-cluster: tries to bring to Linux-Cluster an exclusive shared
storage option, using features of LVM2. I got accustomed to this kind of
volumes when working with Linux-Ha + EVMS solution (using Cluster
Segment Manager plug-in).

When defining a new LVM2 volume four your cluster, you can set it as
cluster-disabled (the volume will behave as a local volume even if it is
on shared storage) or as cluster-enabled (the LVM volume can be
activated on many different cluster nodes at the same time).

Of course, if the filesystem placed into the LVM volume is not a
clustered filesystem (GFS2) a cluster-enabled volume allows a bad
administrator mount a no-clustered filesystem (EXT3) in more than one
node of the cluster which may produce filesystem corruption. This is
because the LVM "open flag" of the filesystem is not propagated through
all the members of the cluster, so there is no knowledge of the state of
the filesystem and this situations can happen.

This can be fixed with some of the options of LVM, specifically the
"enable exclusively flag". This flag, when used over a cluster-enabled
volume, will allow the VolumeGroup to be imported by all the nodes of
the cluster but the LogicalVolumes into the VolumeGroup can only be
activated by a single node. So, only one node of your cluster will have
the LogicalVolume device (for example /dev/VolGrp01/LogVol01) and the
problem explained above cannot happen. This is not about propagating the
"open flag" through the nodes, this is about making the LogicalVolume be
in only one node.

I have developed lvm-cluster as a resource to be used into a service of
your cluster. In the arguments you an specify the name of the
VolumeGroup and the LogicalVolume to handle.



So, I would like to receive the instructions to submit this two
resources to the project to improve them, test them and find any bugs
that could still be in the code. I have made some testing but of course
they need much more to allow them be put into the main project.

Sincerely yours,

-- 
Rafael Micó Miranda





More information about the Cluster-devel mailing list