Right-Size Your Virtual Platform

In virtualization circles, consolidation ratios have become the stuff of bragging rights. A year ago, if you bought a top-of-the-line server and loaded it with memory and connectivity, you might have been able to consolidate 30 or more physical servers onto one trickedout machine.According to Moore's law,what was last year's trickedout server should now be mainstream technology.Does that mean everyone should shoot for consolidation ratios of 30 or more? No, and here's why.

Read the rest of the digital issue on Next Generation Data Centers. [registration required]

To consolidate at such high ratios, most of us would choose the newest four-socket systems.While quite powerful, these servers generally cost more than double their two-socket counterparts. Choosing between the two-socket or four-socket host server is a critical decision for consolidating at modest to very high ratios.The two-socket system can be quite powerful in today's configuration. AMD's and Intel's current systems are built with virtualization awareness, and the performance shows.

What Moore's law does for processor performance, it also does for memory capacity and cost. A few years ago the cost to get server memory north of 64 GB was substantial. Today, 64, 96, and 128 GB are quite affordable.As a result,more, smaller hosts can be provisioned with a lot of memory without breaking the bank.That argues for deploying more two-socket systems with lots of memory in them and not relying on techniques like overcommitted memory.

Whatever your consolidation ratio, the physical server is now a highly critical system and had better be highly available.Virtualization implementations primarily employ one of two high-availability technologies to protect workloads: host-based or virtual machine-based with Microsoft Hyper-V and VMware's Fault Tolerance.For host-based schemes such as VMware's HA feature, there's a reserved inventory of host memory and compute resources known as admission control. This capacity is set aside to handle a host failure by reserving one host's equivalent resources in the existing cluster.VMware doesn't just set aside an idle host, but rather distributes enough headroom across hosts in the cluster to accommodate the entire workload should one of the other servers keel over.

The VMware HA feature is licensed, so there's an associated cost for protecting each host. If you have more modest consolidation ratios with relatively smaller host systems, the host inventory of memory reserved with admission control goes down and saves on licensing costs. To look at it a different way,would you rather HA consumed four CPU licenses or just two?

Then there's the I/O load. Every virtualization environment will eventually come across a few high-performance workloads that will push the storage I/O capabilities. By provisioning less VMs per server, we reduce the risk of network contention and allow more flexibility in where workloads reside.

Here's the bottom line: As virtualization technology matures and servers continue to pack more memory and processing power,what once required high-end systems can now be done with commodity servers. Using two-socket servers loaded with memory offers a great, low-cost consolidation platform and provides flexibility for addressing VMs with both high I/O and high-availability needs.This is yet another case that proves bigger isn't necessarily better.

Rick Vanover (VCP, MCITP, MCSA) is an IT infrastructure manager at Alliance Data. He has years of IT experience and currently focuses on virtualization,Windows-based server administration, and system hardware.Write to us at [email protected].