VMware users pack 16 VMs on each server, despite memory constraints
- 16 December, 2010 08:25
VMware customers run an average of 12.5 virtual machines on each physical server, but high levels of consolidation are driven in large part by proliferation of multi-core servers, and memory limitations may be preventing further progress.
The data comes from virtualization management vendor VKernel, which analyzed 550,000 virtual machines across 2,500 deployments over a period of eight months. VKernel says the data was collected anonymously through free software tools the vendor offers to VMware customers.
The average customer in the sample has 18 physical hosts and 225 virtual machines, or 12.5 per server. Smaller customers are actually driving the ratios up, as businesses that installed VMware on fewer than 10 physical hosts are getting 20 VMs onto each server. Those with 25 to 100 hosts only deploy 11 VMs per server.
When asked why larger deployments would have fewer VMs per physical host, VKernel chief marketing officer Bryan Semple says, "Either the larger environments have an embarrassment of wealth and hence are not as efficient, or they have larger applications running on them."
Putting more than a dozen virtual machines on each box greatly changes the economics of IT, compared to the old model of one application and operating system per server.
While VMware's virtualization software makes it possible, Semple notes that the consolidation ratios also could not be achieved without advancements in multi-core technology from Intel and AMD. With the average customer using a dual-socket, quad-core machine (for a total of eight cores), customers are getting fewer than two virtual machines onto each core.
"It's not that we've gotten any better with capacity planning," Semple says of the rise in VMs per host. "It's that Intel and AMD have succeeded in getting more cores on each host." Two virtual machines per core "is not a huge amount of compression," he says. Semple predicts that over time the number of virtual machines per core will stay the same but that "more and bigger cores will be thrown at the problem."
Since this is the first time VKernel has run what it calls the "Virtualization Management Index," we can't compare today's results with those from several years ago. VKernel's analysis also did not look at Microsoft's Hyper-V, although Semple said the company may do so in the future.
VKernel also examined CPU and memory utilization in the virtualized data centers. Virtualization can effectively let IT allocate more memory and CPU to applications than physically exists in the data center, because resources can be shared across a pool with applications usually not using their entire allotment.
But while organizations are allocating more than twice as much physical CPU capacity as they actually own, they are only allocating about 70 per cent of memory capacity. Businesses may have good reasons for under-allocating memory - for example, they might want a buffer in case of failover situations, Semple says.
But the 70 per cent figure could also indicate inefficiency, he says. Businesses may be buying more CPU than memory because memory costs more up front, but that may be short sighted if extra memory and better planning can achieve higher consolidation ratios and ROI, he says.
VKernel's conclusions have to be taken with at least a small grain of salt, because the vendor sells products that improve utilization of virtual server deployments. But Semple points to data showing that high-density deployments can over-allocate memory without running into performance problems.
Overall, VKernel says the three-year cost is $1,500 per virtual machine for average customers, and only $1,000 per virtual machine for high-density customers.
Read more about data center in Network World's Data Center section.