On 02/23/2012 10:50 AM, hai wu wrote:
There would be some overhead by putting apps into guest, by keeping
app in the bare metal world, it should be faster .. I assume we could
still do the same cgroup configuration in that case, dedicating more
vcpus to the app as well?

On 2/23/12, Gianluca Cecchi<[email protected]>  wrote:
On Thu, Feb 23, 2012 at 3:50 PM, hai wu<[email protected]>  wrote:
Yes, but this way the guests app needs to run on top of the same OS
release, which is not good..
Just to understand the contest, what is the limitation of  not running
"such as cpu intensive puppet app" in a dedicated guest too then?
Eventually dedicating more vcpus to it?
Lack of certification in virtual environments or what?


The overhead of the hyper-visor piece is very minimal for almost everything.

Graphics and real time i/o are still the big exceptions, but work is getting there to offer all types of PCI devices directly to the VM, but not just yet.

Use dedicated drives or other storage space for VMs, pin them to CPUs, etc as has been noted.

Here is your major issue, and one that cannot be ignored:

What happens when you need to upgrade hardware (CPU/RAM). You are looking at some serious hours to move the OS to new hardware.

Virtualize everything, and divorce all apps from the hardware. Need more? Bring up the new hardware and live migrate to it. No downtime at all. This justification alone eliminates almost all possible applications from needing bare Iron. Just use vitrio drivers, and you are always good to go (assuming intel->intel and AMD->AMD).

You may also want to divorce all apps from specific storage requirements. This is harder, but should be in the plans for any large scale roll out of new servers. Think ISCSI, or Fibre Channel, or other reliable NAS before any consideration of local disks. A good storage array can way outperform a local disk. A 4GB RAM stick is cheaper than a hard drive dedicated to swap, so don't swap any more.

And KVM/qemu is not going to force a need for swap on the server without you being able to calculate or see the need in advance.

We are now heading back to the days of centralized services, but in many better ways. The concept is this: Migrate anything from anywhere to anywhere. For a good percentage of apps, we are there now.

Good Luck!

_______________________________________________
rhelv6-list mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/rhelv6-list

Reply via email to