Nathan Scott wrote:
On Fri, 2009-02-13 at 18:42 -0500, Martin Hicks wrote:
On Sat, Feb 14, 2009 at 10:18:47AM +1100, Nathan Scott wrote:
On Fri, 2009-02-13 at 14:29 -0600, Martin Hicks wrote:
I've created my own tree on oss. I'll be pushing patches here
before I commit them to the real git tree, to give the community
time to review.
Good stuff... will have a look soon. At first glance, I was looking
for a pmdacluster / pmclusterd man page ... do those exist anywhere
yet?
Unfortunately not.
I'll write up a couple of man pages, once I've finished my resume ;-)
OK, that'd help in trying to understand this beast. Maybe a
diagram or two as well, after having sniffed around the code
briefly. :)
There is an internal (to SGI) presentation related to this;
perhaps Martin could convince his management to publish it
under pcp-web. Basically, the cluster PMDA does instance
domain aggregation for two or more client hosts running
pmclusterd, which runs in local context mode.
I've left the initscript for pmclusterd (which is the client "push"
daemon) out of /etc/init.d, because it is fairly rare that anyone will
use it.
pmclusterd could be installed by default but definitely should
be chkconfig off by default. The cluster PMDA can be installed
by default (but not configured or run by default, as in ./Install).
It probably also needs some loving to be useful on a machine other than
Altix ICE.
It will work correctly on generic clusters, but will require
some manual config. Compute nodes and head nodes on ICE clusters
have topologically generated hostnames, and hence pmclusterd and
the cluster PMDA can be auto-configured. Basically, pmclusterd
on each compute node needs to know the hostname of it's head node,
which is where the cluster PMDA is running - on ICE, the head
node's host name is guessable from the compute node's hostname.
Hmm. If its a rare / non-generic solution we should probably
not install it by default? Hard to make the call without some
docs on what it is trying to achieve though.
I'm particularly curious about the patch below that I got from
Michael Newton. I presume it is a real bugfix, but I don't
understand this area of the code yet...
That patch looks good to me, BTW.
Yes I reviewed that patch back in December - it's
a bit hairy but I think it's fine.
Cheers
-- Mark
|