At this point even the initial skeptics have accepted that the industry is moving from physical to virtual running on hypervisors such as Microsoft’s Hyper-V and VMware’s vCenter 4.x. It’s also no surprise that many companies are taking this opportunity to refresh their server hardware to take advantage of this trend and reap the benefits of having a private cloud in their environment. There are, however, some simple design considerations to take into account for while doing this. One of which as an IT administrator is choosing a server platform for your environment that fits your needs. Even on something as familiar as a server there may be misconceptions as to what is needed to run an environment in the virtual world.
One of the two most common misconceptions I see when I do virtual sizing for customers (other will be a later post) is that they tend to go with the largest, most powerful server they can buy with 4 sockets and massive amounts of RAM and NICs. This seems like a good approach in some ways as you are able to run a massive virtual pool of servers on the same box (part of the ROI on virtualization right?). While we do have the ability to do this with 4 socket servers like the R910 / R810 / R815 series (latter two in just 2u form factors) and scaling up to 1TB of memory in a single server, this isn’t always the best idea. In fact this is counterproductive to virtual environment in most cases!
Take two possible scenarios where a customer chooses to go with 8 socket license of a hypervisor (commonly sold as acceleration kits) for their virtual environment, there are really two ways to do this; either you go with two extremely powerful 4 socket servers as mentioned above, or you can break it down to four 2 socket servers (Dell versions would be the R710 / R715 / R610) giving you the same number of U space in the rack and same total socket count. In the first option it would probably be something like 2 servers each with 192GB of ram in each server along with high end processors, and around 16 ports of network connectivity. The second scenario typically would be 4 servers with 2 sockets each, 96GB of RAM, and 8 ports of network connectivity for the LAN/SAN/etc. From a cost perspective the licensing portion is still the exact same cost since the sockets are the same quantity, but let’s think about the resiliency of the solutions.
If I were to need to bring a physical server down for any reason (firmware, hypervisor update, reboots, etc), in a two server scenario mentioned above with the 4-socket servers I am effective running my environment with 50% of the total resources that I had before, and have created a single point of failure within my environment should anything happen to that one server! This is something that is avoided with going with the scale out approach of the second scenario above. If I were to bring down a single server in a 4 server environment I am running on 75% of my resources, and I would be able to sustain a second (and even third) server failure and run my mission critical virtual machines while bringing the other boxes online. This makes anything like a software update on the hypervisor that requires a reboot a nearly painless (and comfortable) process. Not only this but as you begin to scale your virtual environment you are able to do it in smaller cost increments with the latest and greatest server technology in smaller two socket units, and since it’s all managed from the same interface within the virtual environment there is little concern with server sprawl.
So in summary, when you are looking at going virtual and choosing the server platform, definitely take the time to research the benefits of scaling out instead of scaling up when building your server and storage hardware architecture, you may be able to build a more robust environment in doing so.
Please feel free to comment here or contact me me here.