On 06/03/2011 04:04 PM, Kevin Saenz wrote:
A lot of those questions you have raised are usually answered in documentation
for the virtual environment you are running.
The docs don't seem to talk about "optimum" setups or have reference
numbers for different configurations of the same hardware.
I spose the tens of thousands of $ you pay for vmware is good for
something afterall ;-P
Let me add a disclaimer that I have ceased supporting organisations that are
smaller than 2000 staff. So my head may be in the clouds
The issue you have is you limit your redundancy if you run RAID 1, and you will
have only 1 TB of space. What VMs are you going to run? If it's a mail server,
SQL and fileserver I would hope that the client is small enough to deal grow
into 1TB shared disk space within the next 3 years and not out grow it in 12
months. In my personal experience I have not seen a small business use less
than 1 TB for files, this was back in 2004-2005
1TB is sufficient, they have ~200gb worth of files atm, and the growth
rate is slow, 2Tb disks are off the shelf items now so i'm safe in that
sense.
My original plan was RAID1 with a drbd mirror of the lot, so in theory
that could handle 3 disks failing and only show a minimal performance
impact.
Now I wonder about the performance implications of that setup. I'm
noticing that iowait is higher in a bursty manner than i was expecting
on their existing setup which is similar but with whitebox hardware
rather than a real "server".
How do you intend support the clustered environment? Granted I have not played
with Xen virtualisation but I would believe that similar rules apply in VMWare
or Microsoft and Citrix. I am making assumptions here: If you are employing HA
clustering My question is how will share disks out across nodes work when your
VM reside on node1's disk and for some stupid reason node1 is dead how will
node2 access the VMs?
The whole point of drbd is to replicate the hdd's across the network,
think of it as raid1 over lan.
So node 1 and node 2 both have the same disk content in their storage
pools. Instances of the VM's can then be started on either node.
ganeti takes care of setting up drbd and such so i can tell it i want a
new instance, i want it to primarily run on node A and it'll
automagically start itself up on node B if node A dies or if i fail node
A it'll live migrate everything over to B so i can start mucking with A
without bothering the users too much.
It really does take care of allot of complexity in neat ways.
BTW, i'm using KVM not xen, kvm feels more linuxy and seems better
integrated. (the makers being bought out by redhat and being part of the
kernel and all)
It was also a 5 line install that just worked out of the box a few years
ago vs Xen which still seems to be a pain in the butt to get working.
I wouldn't worry too much about striping v's mirroring until you consider what
type of cluster you are going to design. To have real HA you might need to
reconsider the idea of supporting any local disks as you will not have real HA
in your design until the disks are external to both cluster nodes. This is big
money but you get what you pay for. for a small home/office for my personal use
I have purchased a DRoboPro FS fully populated with 2 TB drives with a total
capacity of 10TB storage for $6,000 the system is self healing, you can
intermix drive sizes and you use cheap SATA drives, I use it as my test
environment which I run VMs through vSphere, and I have True HA until I lose my
disk subsystem.
I have no hardware SPF's outside of the switch and power, those are a
conscious choice as those components have shown themselves to have
sufficient reliability that the expense and complexity of supporting HA
for them it isn't worth it. (they will have a UPS, but if the power goes
out, all the desktops turn off anyway so its not a big deal)
I was mainly curious about performance vs reliability and stability.
--
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html
