Advanced HA Configuration Help

[dtracey]

Hi guys,

A project I'm working on has some strong application requirements that affect my HA design.  In a nutshell they are:

1) a 3rd party monitoring tool must know where applications (implemented as vApps) are hosted at any one time.

2) applications must not share computing resources.

The bottom line is i need to use a specified failover host as my HA admission control policy, and host affinity rules for my vApps.  Now the difficulty comes when my specified failover host is used or not available.

The default behavior for vSphere 4.1, if the specified failover host is unavailable would be to restart the VMs on any other hosts in the cluster, starting with the host with the most unreserved capacity and vm compatibility first.  This is not desired behaviour as it violates security and control policy for my client.

Now, testing has shown that if i implement a DRS rule for the group of VMs that make up my vApp that they SHOULD run' only on their 'home' host or the failover host, the behaviour is the same.  If i use the 'MUST run' rule, then they become orphaned once the failover blade is lost.

Can anyone shed offer any new ideas about how i can cleverly use HA and DRS rules so that if my dedicated failover host is lost (or i lose an additional host), that the VMs register, BUT DO NOT START on other hosts?

Thanks in advance,

Dan  

[avlieshout]

Did you try the MUST NOT RUN ON HOST rule and configure all other servers.

If the restrictions are that specific, why don't you create a seperate cluster for the vApps using only two hosts?

I did something similar to solve an Oracle WebLogic Licensing issue.

[Gleed]

Can you share details on why there is concern with VMs running together on the same host?  Each VM is secure and Shares can be used to manage resources.

One idea would be to setup an alarm so that anytime the failover host is used/lost a script gets called to disable HA.  The admin would need to intervene to "reset" things anytime there is a failure.

Not a very good HA solution.

Regards,

-Kyle

[dtracey]

Hi Guys,

Thanks for your replies.

In answer to the 'separate cluster for vApps' suggestion, the entire cluster consists of sets of vApps - approximately 10-14 of them across 14 blades.  Some vApps span blades.

The reasons that the vApps cannot share hosts are:

1) App vendors have developed on legacy hardware - and expect isolation.

2) vApps generally consist of 1 master VM and a set of slaves that PXE boot from it - therefore the apps wont run 'properly' if split across hosts (the packet loss tolerance is very, very low)

3) The apps are multicast and use real time audio/video so therefore cannot tolerate packet loss >1 in 1M.  We've tested this and it doesnt run.

I realise that I could use a script to shut these VMs down if they start on a 'wrong' host after two blade failuers, but there is still a loss of control (split vApps) and i didnt want to re-engineer the wheel if a solution exists in the product.

Thanks for your time,

Dan

[depping]

I remember this question from one of our SEs and basically what you are trying to achieve is not what HA is designed for currently. There are a couple of things here:

1) HA will not violate a MUST rule ever. So if you specify a failover host and it and the "source" host fail (and these two belong to the MUST rule) the VM will not be restarted (and not reregistered) as there are no compatible hosts to use.

2) I guess the question and the discussion to have is how Availability rates vs Policy.

Sorry I can't really help you, I think you are too much constraint by business/technical requirements.

Duncan (VCDX)

Available now on Amazon: vSphere 4.1 HA and DRS technical deepdive

[depping]

how many hosts / how many vApps / how many VMs are we talking about by the way?

[dtracey]

Thanks for taking the time to reply guys.

Duncan - in answer to your last question - i'll reply to your PM.

Cheers.

Dan

[DSTAVERT]

I would make sure to post back here anything you may come up with privately. It will help if others can see any resolution.