So far the latter seems to be the prevailing trend as the majority of public cloud platforms and private cloud software solutions start with the foundation of server virtualization. The bare metal options are being positioned more for two purposes:
Auto-provisioning new nodes ofthe cloud – bare metal installation of the cloud solution and the hypervisor
New compute resource types inthe cloud – using new automation capabilities to add a complete physical server to a customer’s cloud tenancy, as if it were just another virtual machine.
The former approach really doesn’t change the cloud value proposition as it is more of an efficiency gain for a cloud administrator. The latter is more interesting as it gives customers of cloud platforms new deployment options that arguably have very different characteristics:
Higher Performance – Clearly if your applications are performance-sensitive, the elimination of the hypervisor would give you the full, raw CPU and memory performance of the server (although it’s arguable how much of a performance hit modern hypervisors insert).
Higher Quality of Service – A non-virtualized resource isn’t subject to noisy-neighbor problems that plague some of the most popular public clouds (although you would still be subject to LAN congestion and network storage sharing issues).
Isolation & Greater Security Configuration – A fully-dedicated server presumably is more secure as you don’t have to worry about hypervisor vulnerabilities (although this concern is also overblown). And the disk inside the physical server is yours and yours alone (assuming there is one).
Better suited to older and more complex workloads – Perhaps one of the biggest benefits of bare metal is that you can now put applications into the cloud that couldn’t (or weren’t suitable) to be virtualized.
However, there are downsides as well. First, do you really need your own server? Is your application going to fully consume all the resources to justify the higher cost? You also lose the option of live migration in the event of a server failure.
From the cloud administrator’s point of view, this is the costliest compute resource they can provide because they lose the ability to allocate the unused resources to other clients and maximize revenue per rack through customer instance churn. And typically bare metal provisioning takes more time and since the full server will be dedicated to the customer, the service provider wants a long term commitment to that resource. IBM SoftLayer, for example, who has been offering bare metal resources in its “cloud” since 2009, prices by the hour but requires a minimum one-month commit to that resource.
Well, at least a few hosters think they can overcome these challenges. BareMetalCloud.com, in 2012, was the first to offer fully bare metal compute resources priced on an hourly basis. Now Bigstep, a spin out of long time traditional hosting company Hostway, is joining them with a fully bare metal Infrastructure as a Service (IaaS) cloud platform targeting high performance workloads. Rather than generic infrastructure, Bigstep is outfitting its cloud specifically for performance – HP rackmount servers, SSD-based iSCSI storage, 44Gbps of Ethernet from each server (four 1GbE ports and 4 10GbE ports per server) and no virtual switching. Now, before you cry cloud-washing, note that the company said its cloud is fully self-service (the end customer controls all provisioning and infrastructure configuration via its self-service portal and RESTful APIs). Bigstep does not require an up-front financial commitment, and its bare metal servers are priced at pennies per hour with a minimum commit of an hour. It further claims that its automated provisioning software can provide requested resources in less than five minutes.
The service launches this week from a Hostway data center in the United Kingdom; the company said it will expand to multiple data centers based on customer need. BareMetalCloud.com hosts out of two Internap facilities in the United States.
Bigstep claims the ability to clone and snapshot server images in minutes, as all OS images are stored on and boot from SSD iSCSI volumes (not from the disk inside the server – that’s reserved for scratch space). It further states that its automatic provisioning software can accommodate requests for resources that are physically close to each other (all three tiers of a web site in the same rack, all your HPC nodes in the two adjacent racks) and contiguously configured (same storage, same switches, etc.).
If Bigstep turns out to match the hype it might just position itself effectively as a niche cloud platform well suited in complement to a general purpose cloud, such as AWS or Microsoft Windows Azure.
However, Bigstep will still face the business model deficiency that comes through physical isolation. Hypervisors let administrators squeeze as many VMs as they can onto the same physical server, driving up utilization and operational efficiency. Perhaps microservers will change this game. As Forrester analyst Rich Fichera points out, server density may soon go up dramatically through forthcoming ARM-based servers such as HP’s Moonshot that reduce physical server footprint by 75 percent. There are power efficiency gains to be had here as well but the overall performance of a microserver might push this offering away from a performance-centric value proposition like Bigstep’s.
While the use of hypervisors is common among IaaS clouds, it’s certainly not a pre-requisite to deliver cloud value (few SaaS and PaaS offerings use it). However, it’s unlikely that a fully bare metal cloud will replace a hypervisor-based IaaS service anytime soon, but there is little doubt that bare metal as an option will become increasingly appealing, available and affordable.