Cloud services, particularly infrastructure- and platform-as-a-service, are well established, but in some cases customers demand more — more control, more access to hardware, more performance, and the ability to pick their own operating environment.
In those cases, they are looking to bare-metal services, a niche that is growing quickly.
As the name implies, bare metal means no software, just CPUs, memory, and storage. Customers provide all of the software from the operating system on up. That means a dedicated CPU, full access to the hardware, and freedom to run custom operating systems.
According to a 2016 Markets and Markets report, the bare-metal cloud market is expected to grow from $871.8 million in 2016 to $4.7 billion in 2021, at an estimated compound annual growth rate of 40.1 percent.
One reason it’s still a small market is because the major providers haven’t really embraced it. Amazon is only just now starting to dip its toe in the waters with a single configuration. Microsoft, remarkably, does offer a bare-metal VMware environment but did it without the blessing or partnership of VMware. Google has no significant bare-metal offering.
IBM is the biggest of the cloud providers with an offering, through the data-center provider SoftLayer, which it acquired in 2013. Other major names supporting bare metal include Oracle and Rackspace, and just recently Chinese giant Alibaba announced plans for bare-metal support. After that, it mostly falls to tier 2 and regional providers.
But Deepak Mohan, research director for public cloud infrastructure as a service for IDC, believes that will change shortly.
“It seems like it’s a fresh wave of bare metal coming to market,” he said. “I assume everybody will have a bare-metal offering by the end of the year.”
On-prem vs. bare metal
Bare metal appeals to customers with on-premises apps that they would like to migrate out of their data centers. Jay Jubran, director of compute offering management for IBM Cloud, notes that most of IBM’s enterprise customers want to move existing apps and outsource existing apps to the cloud and continue to operate as they would on-premises using the same tools. The apps run unchanged but in IBM’s data center instead of the customer’s.
“Bare metal offers a step into the right direction, where you don’t have to change workflows and processes. And it gives you access to an environment hosted somewhere else, so you can start adopting cloud techniques,” he said.
It’s no coincidence that IBM and Oracle are two big proponents of bare-metal provisioning. They are two of the largest enterprise software companies in the world and are feeling the pinch of the cloud. One reason is that their licenses don’t allow their software to be run in a multitenant environment where an unlicensed customer's workloads wind up on the same server as IBM and Oracle workloads that are properly licensed customer.
With bare metal, organizations can move IBM DB2 or Oracle 12 to the IBM or Oracle clouds, no modifications needed, and run those big databases or other enterprise apps on IBM or Oracle’s servers rather than their own.
“If you are running a software package on-prem, you can move to the cloud with no performance tradeoffs,” said Mohan. “This becomes increasingly relevant when you go to software packages that have the necessary performance to take advantage of the performance benefits of bare metal, like analytics and high-end databases.”
Addressing cloud performance issues
Beyond app migration, there are three major selling points for bare metal provisioning: performance, hardware access, and customization. In a virtual environment, you share CPU and memory space with other customers, and performance can often vary from one instance to another.
Performance can be an issue because of a phenomenon known as the noisy neighbor. When customers launch a virtual machine in a public cloud, they have no idea what workloads other customers might be running on the same hardware, competing for CPU cycles and memory.
This uncertainty leads to an interesting practice where some Amazon EC2 customers will batch-launch one hundred or more virtual machines, test them with the identical workload, and kill the instances that don’t perform well.
“It’s possible to have a scenario where you run a workload once, then run it again at another time and get half the speed as the last instance,” said Zach Smith, CEO of Packet, a bare metal provider. “You don’t have that in bare metal environment. You have consistency.”
This issue plagued Packet customer NS1, a startup that provides DNS and traffic-management services for load-balancing between data centers — services that are very sensitive to latency and performance in general.
“We were running into issues with noisy neighbors and couldn’t guarantee performance at all times due to other things happening on the host,” said Nathanael Jean-Francois, senior network architect at NS1. “Performance was all over the map, which made it tough. That’s why we went where we were the only tenant and can run it the way we need to."
He said performance could vary from 5 percent to 20 percent day to day, and a bare-metal service solved the problem.
A second selling point for bare metal is access to hardware. Virtualized environments limit access to hardware such as the NIC. But customers might want to access the NIC, for example, to accelerate certain packets using DPDK, a set of libraries for fast-packet processing. In a virtualized environment, if there is such access, it is highly restricted and virtualized.
Customized options are OK
Finally, bare metal allows running custom environments. In EC2 or Google Compute Engine, all that's available is their brew of Linux without the option of making customer modifications.
Downside of bare metal
The drawback to bare metal is lack of software support. In an Amazon EC2 instance, a good deal of the work is already done. Customers just upload their apps and code. With bare metal, provisioning the hardware falls to the customer, not the provider. As Jean-Francois put it, “A lot of the things you have to build yourself, but it’s not impossible to do.”
Packet’s Smith believes bare-metal workloads will get bigger and more specialized in the near future.
“I think a year from now there will be more metal and more offloading to GPUs, smart NICs and FPGAs. There will be more and more specialized hardware, and that will lead more people to bare metal,” he said.
Mohan said the move to bare metal won’t be as fast as it has been to virtual environments.
“These are custom apps and it will take longer to move them to the cloud. But even if it does take a six- to 12-month lead time, we expect beyond that it will be a strong preferred path for anyone who wants to load a service in the cloud,” he said.
But it will happen.
“Finally we’ve reached that point that benefits from these services are getting acknowledged by the market. These higher end companies moving to the cloud want more customization for what they run,” he said.