Quick Vs. Live Migration, Redux
Greg Shields is our "Virtual Architect" columnist, and he's awesome. His column on virtualization strategies for disaster recovery, coming in the May/June issue, is outstanding, and a must read. (Not a subscriber? Do something
Greg is also an author for Realtime, an online publishing company. He recently responded to a blog posting of mine that referenced a demo VMware published about Quick Migration, and how it drops TCP connections. VMware's point is that Quick Migration is inferior to Live Migration, which can move VMs from physical server to physical server without dropping connections.
Greg's key points:
"But, what is important about either Quick Migration or VMotion is that these sorts of hot migrations are typically reserved for one of two functions: Planned downtime and load balancing. With planned downtime, an extended outage like the one Keith discusses is likely to only occur during standard outage windows when users and their clients are less likely to be using the servers. Thus, Quick Migration's added delay shouldn't necessarily impact operations in this case.
The other case, load balancing, is harder. ESX when in fully automated mode can rearrange its load constantly throughout the day. In a Quick Migration world with extended outages like we currently see, this simply isn't going to work from an operations standpoint."
In a private e-mail, I took issue with Greg's assertion about planned downtime. Isn't the ability to plan your downtime during regular work hours, with full staff on hand, better than doing it at 3 a.m.? Greg responded, and I think what he wrote is thought-provoking (like most of what he writes):
"...The real conclusion of my argument is that you've got to weigh the cost against this benefit. I've never been a fan of "let's implement [insert tech here] because it makes my life easier as an admin". There's never been a great ROI for making admins' lives easier. If it benefits the users, or security, or configuration control, then yes. But if it only "makes the admins' lives easier", then it's not a good play.
At least that's been my recurring opinion.
Right now, VMware's something like $5K to $6K per dual-socket + the multi-thousand dollar cost for VC. Microsoft's effectively free per dual-socket + $500 for SCVMM Next. The affordability vs. return question Its growing harder in my book."
Good points, all. But I would bring up two other objections. First, I've talked with several admins who say that a lack of Live Migration capabilities is an absolute deal-breaker for them, at any price.
The second point is that admins aren't saddled with just two choices. For instance, both Citrix, with XenServer, and Virtual Iron offer Live Migration in their products, and at a significant cost savings over VMware. That's because they use Xen as the base hypervisor, and Xen includes Live Migration. Xen, you may know, is a free, open source hypervisor that has been included in various versions of Linux (both free and commercial) for years now. That means you can get Live Migration for nothing, if you're expert enough to use it.
So, you have choices -- many of them. Which ones are you considering? Let me know.
Posted by Keith Ward on 05/05/2008 at 12:48 PM