Getting Down To Bare Metal On The Cloud
February 9, 2017 Jeffrey Burt
When you think of the public cloud, the tendency is to focus on the big ones, like Amazon Web Services, Microsoft Azure, or Google Cloud Platform. They’re massive, dominating the public cloud skyline with huge datacenters filled with thousands of highly virtualized servers, not to mention virtualized storage and networking. Capacity is divvied up among corporate customers that are increasingly looking to run and store their workloads on someone else’s infrastructure, hardware that they don’t have to set up, deploy, manage or maintain themselves.
But as we’ve talked about before here at The Next Platform, not all workloads run best on virtual systems. This is where bare metal systems running in a cloud come in. For example, Rackspace Hosting has taken to building bare metal systems based on Open Compute Project (OCP) specifications for its OnMetal service and offering the same style of computing as can be found in virtualized environments – easy-in, easy-out, pay-as-you-go, and instantly available.
In 2015, two years after being founded, Packet joined the bare metal cloud space, led by Zachary Smith, who had gotten his first taste of the cloud business as an executive with cloud hosting company Voxel until it was sold to Internap in 2011 for $35 million. Now the company runs four datacenters around the world that offer five different server configurations that can be essentially rented by the hour and are aimed at developers. Speaking with The Next Platform, Smith said he saw a couple of trends in motion that convinced him to get back into the infrastructure game in 2013, and that those trends will drive what he expects will be the rapid deployment of smaller bare metal cloud datacenters throughout the country over the next few years. Packet plans to have a hand in many of those datacenters, either as the operator or in partnership with other companies.
One of the key trends was the shift in who was looking to buy compute capacity. The older developers and IT professionals who were comfortable and interested in the ins and outs of a computer were still buying big systems from OEMs and installing them in their datacenters. However, there were younger developers who had no interest in how things worked and didn’t want to worry about all the security, storage, and I/O hassles that came with compute systems; they just wanted to know that the compute resources worked and their workloads would run. It is a generational change.
“The Millennials are buying servers now,” Smith says. “Why would they possibly want to buy a datacenter and worry about the RAID card?”
The other essential trend came with the introduction of Docker containers. Before containers, moving workloads between datacenters and clouds and between different clouds was difficult. Containers made that easier, and that portability has recently liberated workloads and given developers greater freedom in where they run them, Smith says. They also can be spun up faster than virtual machine instances on hypervisors, and users can pack more containers than virtual machines on a host system, increasing their efficiencies.
All this is driving the need for cloud infrastructure, and the move in that direction is strong. IDC analysts last month predicted spending on cloud IT infrastructure – including servers, storage appliances and networking gear – will hit $44 billion this year, an 18 percent increase, with 61 percent being spent in public clouds. However, Smith says that the big cloud providers like Amazon are simply going up the stack and making everything a homogenized as-a-service.
“We are moving farther and farther away from having the capability to innovate,” the CEO said. “You can consume servers, but you can’t innovate.”
Packet runs 6,000 to 7,000 servers in four datacenters – one each in the New York City, San Jose, Amsterdam, and Tokyo – and currently supports about 7,000 developers running 50,000 deployments a month, with most deployments spinning up within eight minutes. The server lineup includes four configurations powered by Intel processors. One uses its low-power Atom C2550 chip, while the others use the Xeon E3-1240 v3 and the E5-2650 v4 and E5-2640 v3 processors, with pricing ranging from 5 cents an hour to $1.75 an hour. Last year, the company added a 96-core configuration powered by two Cavium ThunderX ARM chips, which can be used for 50 cents an hour.
The servers using these motors, from Dell, Supermicro and Quanta Computer, run a variety of operating systems, from Windows Server 2012 R2 to the popular Linux distros. Adding ARM gave developers more options, and has attracted such workloads as those leveraging Docker’s eponymous containers and Google’s Kubernetes container orchestrator as well as Android workloads.
Packet also offers block storage for 7 cents to 15 cents per GB per month, network services that include a native IPv6 stack, BGP, and Anycast support and up to 20 Gb/sec per host. In addition, customers can use Packet’s technologies to build their own private clouds.
Smith is expecting more workloads to move to bare metal clouds. Platform developers are attracted to the ability use 100 percent of a processor rather than a part of a chip that has been virtually sliced into parts. They’re assured of having all the processor power they need. In the same vein, networking vendors are offered a full L3 topology rather than having to use virtual routers, and users can bring their own L2 overlays. He also pointed to emerging workloads, from virtual reality and analytics to telcos develop virtual-network functions.
In 2017, Smith sees an expansion of the bare metal cloud space, with datacenters cropping up in 50 to 100 different locations across the country to meet a growing demand for cloud facilities closer to where the users are. How that will roll out is unclear, he says, adding that Packet will have a roll in it, either as the operator of the datacenter or a facilitator that will help people find it.
“We can provide the interface,” Smith says. “If we the job right, there will be a lot of places you can get compute through our API. If you need compute in St. Louis, you can either try to find it in St. Louis or use our API to find what you want in St. Louis. And if there’s no place for computer in St. Louis, I will put one there.”
Pricing also may be a focus in the market this year, in such areas as spot markets where extra compute capacity is put out to bid and dynamic pricing that is market-driven. “We are trying to guess what all this is worth, but if it’s dynamic pricing, it’s the market that tells you what it’s worth,” he says.