By Nathan Brookwood, Research Fellow, Insight 64

In the days before virtualization and multi-core processors came to the datacenter, it was easy to determine the optimal specs for your server CPUs. You just benchmarked each application on a range of servers with various CPUs and chose the processor that could comfortably handle the projected peak workload. Most of the time those CPUs were running at less than 15 percent of their maximum performance, but the work got completed and you as the IT manager could sleep soundly most of the time.

Then virtualization became a thing, and the world became more complicated. You could load several applications, each in their own VM, onto a single server and achieve far higher levels of machine utilization, thus making your hardware budget go further, but the VM software was not free. In addition, for each VM, the OS and application had to be licensed at the server level, adding to software licensing fees. Just to make your life more difficult, software vendors each took different approaches to how they charged for VM software and associated applications.

Until last year VMware, the industry leader, charged a flat fee per CPU socket. As CPU suppliers increased the number of cores on their chips, this made servers with high core counts a real bargain for VMware's customers. The folks at VMware, no dummies they, responded by capping the number of cores supported by a single license to 32, so that any socket with more than 32 cores requires an additional license.

Nutanix, the number two supplier, usually bundles its VM software with its hardware, so they get paid more when you buy bigger servers with bigger CPUs and more memory to support them. (Of course, you may still have to pay the VMware piper unless you are using the Nutanix AHV virtualization layer instead of VMware. Nutanix has made great strides with their AHV virtualization layer over the last two years. Customers may want to consider using AHV to further reduce solution licensing costs when deploying Nutanix.)

Microsoft has been charging for most of its software by the CPU core for years, so they don't care how many cores there are in a system, or whether those cores sit on one or several CPU chips. Even so, you might want to consult with a CPA to understand the implications of the Microsoft Enterprise Agreement (EA) for its Azure Stack license and it is worth noting that the MSFT licensing schema varies from application to application.

It's hard enough to keep a datacenter running smoothly under ordinary conditions, but how is an IT manager supposed to cope with all this licensing complexity? To simplify the problem, you can think about it as a game. The virtualization vendors want to extract the maximum fee they can for the operational benefits they supply. The virtualization customers want to minimize the fees they pay for that software. The tactics you use in deploying your virtualized infrastructure may vary, depending on your vendor, but here are a few suggestions:

1. Try to optimize the number of physical cores you use in each system. Thirty-two is a magic number for VMware. If it looks like you need forty cores, explore whether adding more DRAM or going to a 32-core chip with faster clock frequencies can handle the job. The memory and CPU upgrades may well cost less than the extra VMware license fee.
2. In general, opt for faster, rather than a larger number of cores in your server deployments. Many middleware products charge by the core, independent of what the VM supplier charges. The fewer the cores, the less you'll pay those virtualization vendors.
3. If your application just won't fit in a 32-core CPU, explore whether a 48- or 64-core CPU can do the job. If a single CPU just can't get the job done, you may be forced to go to a dual-CPU configuration. You will have to bite the bullet on software fees, but you'll avoid the need for another server with all its associated overhead and space requirements. Keep in mind that once you go beyond 32 cores, you will need to buy two VMware licenses for that processor, whether it is a 40-core Third Generation Intel Xeon SP or a 48- or 64-core Third Generation AMD EPYC processor. Depending on your workload, you are likely to get more bang for your software bucks with the EPYC solutions.
4. And lastly, it is important to evaluate the full impact of VM/OS/Application licensing costs in relationship to core counts before deciding on your CPU. Single socket (CPU) servers may be a viable option. With multiple core count options available, many of the latest single-socket EPYC servers have IO, memory and performance that exceeds many dual socket servers.
   

Regardless of which hardware virtualization software supplier you choose, there has never been a better time to upgrade your IT infrastructure. The amount of processing power you can purchase per dollar spent is absolutely awesome, and the sophisticated software tools available today let you deploy that hardware with less effort than ever before. It's a great time to be an IT manager.

You can learn more of my views by reading my white paper "How HCI Simplifies IT and Lowers Costs" at https://www.amd.com/en/processors/incredible-shrinking-data-center.

Disclosure: While the content in this article is my own, I have a paid relationship with AMD that enabled the creation of this article and the whitepaper, "How HCI Simplifies IT and Lowers Costs."

##

ABOUT THE AUTHOR