Elias Khnaser, Honorary vmLimited Ambassador, Talks About Shared Nothing Live Migration

Back in 2009, Elias Khnaser posted a very badly informed article on InformationWeek with on why you shouldn’t deploy Hyper-V.  I gave it a good bashing, tearing down his points one by one with actual facts.  Well, just in time to be hired by Tad, Elias is back and at it again!

First off, let’s look at the title of the article:

Shared-Nothing Live Migration: Cool, But Not a Game-Changer

Hmm, I have to disagree.  No one else does this right now and it’s a real problem for some.  Think of a large data centre going through a hardware or network refresh.  They can’t afford down time while the export, carry, and import VMs.  They want to be able to move those VMs with the minimum of downtime, and maybe even eliminate downtime.  Shared Nothing Live Migration achieves this.

… entertain this scenario for me: a VM with 1 TB virtual disk … wait, Eli, you are not realistic, you might be thinking … fair enough — what about a VM with 500GB virtual disks? Moving that amount of data over a 1 GB Ethernet or even a 10 GB Ethernet is not quick or feasible in most environments …

Firstly, the vast majority of VMs are small.  And while Windows Server 2012 Hyper-V can support 64 TB VHDX, they will the tiny minority.  And to be honest, not only will these size VMs be few and far between, but I’d expect them to run on clustered hosts with shared storage so Share Nothing Live Migration wouldn’t be needed … except for well planned and scheduled migrations to another part of the data centre.

If you do have lots of 1 TB VMs to move around, then you’re a very large data centre and you’ll have lots of budget for big networking such as Infiniband with RDMA to speed things along.  For the rest of us:

• DCB and converged fabrics
• Everything from 1 GbE to Infiniband
• RDMA
• QoS
• Many ways to architect our networking based on our needs

In my opinion, the market that will make most use of Shared Nothing Live Migration are the public clouds or hosting companies.  To keep costs down, many of them are using non-clustered hosts.  And from time to time, they want to replace hardware … a planned operation.  They can do this with this new Hyper-V feature.  And I can speak from experience: most hosted VMs are very small and would cause no issue to move, even over a 1 GbE network.

… this would only be used as a maintenance technique which is what it was slated for anyway.

Of course!  As a virtualisation expert, Elias, I’d trust that you know the difference between high availability (HA – normally reactive) and LIve Migration (LM – normally proactive and planned).  But if one is stuck with working with vSphere standard, then one probably never gets to implement these features because of their high vTax.

To think that you will constantly be copying large virtual disks between hosts is not practical and is not scalable.

Really?  Elias have you even looked at why Shared Nothing Live Migration exists?  It’s not there for load balancing or as an alternative to Failover Clustering HA.  It is there to allow us to do strategic moves of virtual workloads from one cluster to another (or a standalone host), one standalone host to another (or a cluster), from one part of the data centre to another, or even from a private cloud to a public one.  If you’re doing this all of the time with all of your VMs then you need to take a long look at yourself and your planning.

… you cannot use high availability with this feature, and that makes sense since you need a point of reference for HA to work. How can you recover a VM when its files are on the host that failed?

You’re confused and you’re wrong:

1. You don’t need Shared Nothing Live Migration within a cluster because, strangely enough, there is shared storage in a cluster. With the VMs storage on a CSV, you don’t need to move the VHD(X) from one host to another, or from one storage device to another, to LM a VM from one host to another.
2. You can use Shared Nothing Live Migration to move a workload from/to a cluster.

Live Migration (proactive VM move) is not HA (reactive failover).  Yes, LM has been separated from Failover Clustering, but they are certainly not mutually exclusive.  And anyone who sees Shared Nothing Live Migration as an alternative to HA needs to reconsider their career path.

I expect that when VMware does feature catch up with Windows Server 2012 Hyper-V, Elias might have a change of heart regarding Shared Nothing vMotion   But until then, I expect we need to beware of bogus nastiness and stay vmLimited.