Are you paying too much for virtualized systems either on-premises or in the cloud? A new report by Control Up tracked the performance of over 140,000 servers and found significant amounts of CPU and memory capacity were under-utilized.
It’s not a huge secret that data centers are over-provisioned because it can be expensive to suddenly run short of resources and have to quickly buy and deploy more. But Rotem Agmon, technical lead for virtualization at Control Up, said companies can spend less and still have piece of mind.
Control Up offers analytics and monitoring systems for virtualization. Agmon noted there are other solutions for managing IT resources and said the point of study was to make companies more aware of the issue, not to pitch his own company’s services.
“It is a well-known that over-provisioning takes place in data centers everywhere in the world, we just never had sound facts to base the claim on,” Agmon told eWEEK. “Being an analytics company, we have access to huge amounts of data and that’s what we studied.”
The preliminary report, shared exclusively with eWEEK, covered 148,000 production workloads across 943 organizations worldwide. Control Up inspected the number of CPUs and memory allocated to each server and then compared those numbers to the amount actually used over a 60-day period.
The results showed that almost half of the VDI (Virtual Desktop Infrastructure) machines were given too much RAM, and many general-purpose servers were almost 80 percent over-provisioned. VDI is the process of running a user desktop inside a virtual machine that resides on a server in a data center. Servers were over provisioned more than six times as much as desktop VMs. The report speculates this may be because a desktop VM is for one user while the server is a shared asset that is more critical to keep up and running.
“The study offers some solid food for thought in regards to over-provisioned IT resources,” Pund-IT analyst Charles King told eWEEK. “It’s an interesting issue to consider, especially since virtualizing x86 systems (as well as enterprise-class RISC and EPYC systems) aimed to maximize what were up until that point mostly underutilized system resources.”
King makes a good point that one of the main benefits of virtualization is saving companies money on hardware costs, but Agmon said that doesn’t mean you can’t save even more. He said most IT administers will avoid right-sizing VMs because it’s a very complicated undertaking.
“They usually will accept the architecture design from the vendor and pay much more than they need to,” said Agmon. “I’ve worked for many of the big hardware vendors. When the customer runs out of capacity, they don’t send someone to see if they can make it run more efficiently, they sell you more servers.”
While a more efficient system will help companies save on the cost of RAM, Agmon said that having more CPUs than you need is not just a question of overspending.
He compares the situation to when two people walk into a restaurant. Typically, there are a lot of options as to where they can be seated because there are many tables set to accommodate two diners and there might also be seating at the bar. But when eight people show up without reservations they can expect a long wait because there are fewer tables that can accommodate that many diners.
“Fewer CPUs increases the consolidation ratio and eliminates bottlenecks in your virtualization stack,” he said. “When you over-provision CPUs your VM’s performance suffers because the hypervisor gets more complex.”
As for specific cost-savings, the report states it’s especially difficult to estimate the financial impact of over-provisioning CPUs because they are constantly being shared. On the other hand, estimating the impact of memory allocation is relatively straightforward math. In the Control Up study there was 26.2 GB of RAM allocated at a cost of $69.3 million for 71,845 servers running an OS. Of that total, Control Up estimates 77 percent of the servers were over-provisioned leading to $30.9 million overspent on RAM.
A more complete version of the report due out the end of May will also look at the impact of upgrading your operating system. There has been at least one surprise in the preliminary results.
“I was under the impression the more recent your operating system like Windows 10 the more resources you’ll use. That’s usually the case that the newer OS is a memory and CPU hog,” said Agmon. “But with Windows 10 and Windows 2016 Server we’ve seen massive drops in utilization, they’re super-efficient. Microsoft’s done some amazing work.”
