Rackspace Intros Dedicated Servers That Behave Just Like Cloud VMs

Seeking to make cloud infrastructure performance more predictable for its customers, Rackspace is launching a new “bare-metal” server offering, giving users the ability to spin up and down dedicated servers just like they spin up and down virtual machines in its OpenStack cloud.

Rackspace will charge for the service by the minute, meaning anybody will be able to rent a powerful dedicated server sitting in a Rackspace data center for 20 or 30 minutes at a time. They can request it online and have it up and running in a matter of minutes, using the same OpenStack API and tools used to provision and manage cloud VMs.

Ev Kontsevoy, the company’s director of product, said the problem with cloud VMs was the multi-tenant nature of the service. Your application is always sharing physical server resources and network connections with other applications, making its performance fluctuate constantly.

Currently, companies that start on cloud infrastructure, move into colocation data centers when they start growing and need to scale to avoid this problem. As they move, howver, they lose the elasticity of cloud.

This predicament is what Rackspace is trying to address with its new OnMetal service, offering dedicated machines with all the utility-computing benefits of Infrastructure-as-a-Service, Kontsevoy explained.

Rackspace president Taylor Rhodes announced OnMetal Thursday at the GigaOm Structure conference in San Francisco.

"Faster than SoftLayer," billed by the minute

This is certainly not the first bare-metal cloud service ever launched. SoftLayer, the company IBM bought last year, was founded around this concept, and Rackspace itself has offered bare-metal servers before.

OnMetal is very different from both of the above, Kontsevoy said. It performs better and provisions servers much faster.

The company’s previously existing bare-metal offering is more like traditional dedicated hosting, where a customer has to wait hours or sometimes even days to have a server provisioned. They would be billed by the month.

IBM SoftLayer charges for its bare-metal cloud servers by the hour, but its performance is subpar, Kontsevoy said. The provider uses off-the-shelf Supermicro servers which are not optimized the same way Rackspace’s hardware is optimized, he explained.

“Open Compute freaks”

Using reference designs for Open Compute servers, Rackspace has developed three custom designs specifically for OnMetal. The company has identified four workloads that stand to benefit the most from the bare-metal cloud service (processing web requests, background processing, RAM-based caching and database servers) and modified Open Compute designs for them.

The Open Compute Project is an open source hardware and data center design initiative started and led by Facebook. Open source server designs currently available through OCP were originally developed by Facebook for its own purposes.

Rackspace has designed its own OCP-based servers for its traditional VM cloud offering as well.

The OnMetal server is 100 percent solid state, Kontsevoy said. There are no moving parts, no heat (other than CPU) and no vibration, since there are no spinning disks or fans. Cooling is external.

The servers are memory-heavy: half a terabyte is at the low end of the range. “We call them Open Compute freaks,” he said.

Cheaper than VMs for large workloads

OnMetal is currently in limited availability, with a few customers “kicking tires.” Rackspace expects it to go into general availability in June.

The initial deployment is in the company’s Herndon, Virginia, data center. The company will dedicate more data center capacity to the offering as it becomes necessary.

Rackspace has not announced pricing for OnMetal. Running a large workload on it, however, will cost a lot less than running a large workload on cloud VMs, since that’s what it’s optimized for.

The offering aims to bring together the best of both worlds: the capacity and stability of colocation and the scaling capabilities of cloud.

“Ultimately, this leads to much simpler scaling,” Kontsevoy said. “If you are transitioning from a public cloud to colocation, come and take a look at this.”