|
I didn't say Linux-HA
is perfect. It is not. Other cluster solutions are far superior, such
as MC ServiceGuard, VCS, etc. However, my meaning was that RH Cluster
sux, and is probably one of the worst HA clusters I have had the
pleasure of using.
Ez.
guy keren wrote:
On Mon, 21 Aug 2006, Ez-Aton wrote:
RH Cluster is a bad joke.
linux-ha is also not so good (e.g. it cannot recover from loss of access
to external disks).
I have used various HA solutions, including VCS, SunCluster, HACMP, and
even MSCS, and without a doubt, RH Cluster sux. It lacks features, and
its main defensemechanism against split-brain is "Shoot the Other in
the Head" via its UPS, it's Fibre link, or the likes (they call if
"Fence"). Instead of better logic (how to detect split-brain? How to
prevent it?), they use brute-force in a way I didn't like.
there is no theoretical solution to a real "split brains" situation. most
clustering software use some sort of SCSI reservation to prevent this -
but then, if the split is complete - there'll not be proper access to the
disks anyway (and in a real HA system, you have two sets of disks - so
there can be a split between them as well).
In my simple tests (used HTTPD as a resource) the cluster was unable to
recover from a simple "pkill httpd" on the active node, and completely
flunked my tests.
I would recommend you check Linux-HA. It is looking OK, seems adjustable
to your needs, and would probably work better. It is a bit more
complicated to setup (although it's not too complicated), but it can be
controlled via simple scripts, which can probably do what you wanted it
to do.
albeit linux-ha being better - it is too problematic (and uses the same
STONITH method during split-brains - and ofcourse STONITH can't work when
there is a real communications problem between the two servers).
note that there are some commercial cluster software for linux, which
looks far better, features-wide, when compared to redhat cluster or to
linux-HA.
Ez.
Ira Abramov wrote:
Quoting Vitaly Karasik, from the post of Sun, 20 Aug:
so, is there a config error here, or should I dump the whole iSCSI
concept? is there a way to install a red-hat cluster of three
CENTOS3 machines with no common storage? I just need IP addresses
and processes moving around between the nodes, the application
vendor ONLY supports Red Hat 3 and its clustering, but won't supply
instructions or recommended procedures. arrrrggh!
As far as I remember, RHEL3 Cluster Manager cannot work without shared
storage anddoesn't support iSCSI device as a shared storage (at
least, RH doesn't promise that this configuration will work stable)
it works just fine. RHEL Cluster with two common raw devices for the
quorum, I didn't bother setting up GFS atthe end, since it was not
important.
I was very disappointed from the RH cluster manager though. all it does
it move a list of services without dependency on eachother. it's quite a
lot but it's missing some needed features, like defining a logical link
or block - service A and B must migrate to new nodes together, but not
to one that already runs service C for instance. nope, I can only define
to which nodes each service migrates and that's it. For instance, y
client wanted a very simple case where three machines run two services.
if any of the three machines fails, the other two take over the two
services that need to run, but I can't have both services migrating to
the same node, and now I cannot prevent this using this tool, I'll have
to make funny improvizations in the startup files to get it to "fail"
for the cluster manager and force it to migrate it further to another
node if this one is busy. this is an ugly kludge, and the only "right"
solutiong, per RHEL, is to have 4 rather than 3 machines, each pair
takes care of one service and that it. rediculous :-(
|
