This may be a side note but here is my take on the whole thing:
Our cluster has 132 nodes, plus we have another cluster with 60. I am putting Gentoo on both of them now. (Tonight I am doing the big cluster)

Update procedure:
Compile the new update using a compute node and the scheduler:
# qsub -b y -N update emerge -B <foobar>

# Now install.
pdsh -a
pdsh> emerge -k <foobar>

Now every node has the update. =)

Best part about it is that we can setup pdsh to connect to our install image as well as our compute nodes so everything is updated all at once.

The only thing we had to do to get this to work was NFS mount the / usr/portage directory.

Plus if a node gets out of whack with everything else we run:
# ssh <node> dd if/dev/zero of=/dev/<hda,sda> bs=1024 count=1
# ssh <node> reboot

After that YACI takes over and everything is imaged in about 10 minutes.


total update process takes about 5 minutes minus compile time.


On Nov 20, 2005, at 5:51 PM, Stéphane Lacasse wrote:

For this reason I way prefer Rocks Cluster
that is really a breeze to install, but they do not have the bootp
paradigm...

Which comes back to my original post that started the thred. I wat to make an entirely _Gentoo_ based cluster. A good reason for this is that Gentoo well...is Gentoo, don't want/need to start the philosophical debate on why Gentoo is better than RHE(WS), CentOS and all on which Rocks is based...

Hey, if you can create a Gentoo base cluster that is as easy to install
and maintain than Rocks Cluster is, you have my support.  Being using
Gentoo for 3 years, the benefit is obvious to me ;)

each node is a full image on it's own, but managmenet is
centralized thrue the headnode.

O_o.... now _that_ is something I would call inefficient. I can't immagine a
1024 node cluster running off 1024 images stored on one server.

In fact, there are no images.  Each node is a full system on the hard
drive.  The head node can "order" the nodes to updates it's softwares,
so updating the headnode will take care to also update the nodes
automaticaly.

Like you said, it's beside the point, but I think you could take a look at their design and draw some inpirations from it. An easy install like
Rock Cluster but instead of using Kickstart(tm) files, would use the
emerge system + distcc + quickpkg.

--
[email protected] mailing list



--
[email protected] mailing list

Reply via email to