1 |
Giacomo Bagnoli <g.bagnoli@××××××.com> 2009-05-18 17:06: |
2 |
> Brian, |
3 |
> first of all thanks for your answer |
4 |
> |
5 |
> Il giorno lun, 18/05/2009 alle 09.29 -0500, Brian Kroth ha scritto: |
6 |
> > I remember looking at this a while ago as well and being fairly |
7 |
> > confused. I'm fairly certain that lvm2 with the use flags you mention |
8 |
> > is the way to go, so you might have to unmask them in your |
9 |
> > /etc/portage/profile/use.mask (add -clvm, -cman). |
10 |
> > |
11 |
> > OCFS2 generally uses it's own different cluster stack than CMAN (which |
12 |
> > uses openais). Since I believe 2.6.26 you can use the other one, but I |
13 |
> > haven't found it to be as stable. |
14 |
> |
15 |
> I didn't understand here what you mean. Which one is "the other one"? |
16 |
> (maybe it's just my english :P ) |
17 |
|
18 |
Since 2.6.26, OCFS2 has two cluster stacks available to it: |
19 |
- ocfs2_stack_o2cb: which is the original stack |
20 |
- ocfs2_stack_user: which is the one that requires cman and openais |
21 |
|
22 |
> > It also requires quite a bit of extra |
23 |
> > configuration for proper fencing, whereas that's built into the OCFS2 |
24 |
> > stack. Since you're doing this with DRBD it sounds like you're trying |
25 |
> > to do it on the cheap so proper fencing probably isn't available |
26 |
> > anyways. |
27 |
> |
28 |
> Yes, you're right. The only way I have to get a sort of STONITH (if it |
29 |
> can be used, I've still to check this) is IPMI which is working well so |
30 |
> far. |
31 |
|
32 |
I believe that will work just fine. |
33 |
|
34 |
> > If you aren't doing anything with LVM that needs to be cluster aware |
35 |
> > (eg: just a concatenation of multiple PVs, no striping) then it should |
36 |
> > be just fine without CLVM. The key is that you /must/ take all but one |
37 |
> > node offline (umount; vgchange -an) before you do any LVM operations |
38 |
> > such as adding more extents, else they'll have an inconsistent view of |
39 |
> > the VGs/LVs. This also isn't technically supported but I've run it |
40 |
> > without problems for a year or so now. |
41 |
> |
42 |
> The problem is that a node will use (mount) some LVs and the other one |
43 |
> some others. So I can't do the fast way as you suggest, as I need to |
44 |
> stop all services that use data on LVs before unmounting. |
45 |
|
46 |
But, with OCFS2 you can (even temporarily) have a single node mount all |
47 |
the LVs, move your service IPs over, and _then_ umount and stop them on |
48 |
the other one while you do your changes and then remount them all. |
49 |
That's the great things amount a clustered fs - more than one node can |
50 |
play in the sandbox without problems. |
51 |
|
52 |
Also, I'm gonna put a plug in for Heartbeat [1] for service management. |
53 |
It's much more powerful than rgmanager in my opinion. If you're dead |
54 |
set on using the openais cluster stack you can also use pacemaker [2]. |
55 |
|
56 |
The model I typically use for Heartbeat, is to have it run a couple of |
57 |
dummy monitoring scripts as resource agents clones that each set an |
58 |
attribute as to the service's health on a machine. Then using the value |
59 |
of that attribute I setup rules to move service IPs to the "healthiest" |
60 |
node, where your definition of "healthiest" is completely dependent upon |
61 |
your monitoring scripts. |
62 |
|
63 |
Cheers, |
64 |
Brian |
65 |
|
66 |
[1] http://www.linux-ha.org/ |
67 |
[2] http://clusterlabs.org/wiki/Main_Page |