Open Compute is an initiative to openly share innovative server, storage, datacentre, and other hardware designs. The key word is 'open', as the designs are open source, meaning any organization can take the plans and have their own hardware built to those specifications, or even alter the designs to better match their specific requirements. Those changes can be contributed back to the community ' just like open source software such as Linux or OpenStack ' and the community can further innovate by building on those contributions.
If you read enough articles about startups, eventually you come across the term 'disruption' or 'disruptive innovation.' The idea behind disruptive innovation is to find markets that, for various reasons, are stagnant or unwilling to change.
Businesses in these markets tend to make a lot of money off of their high-end customers and are comfortable continuing to do so ' and sometimes, so are their customers. Thus, there is little to no desire in these industries to change or innovate. But, if a startup company ' or in this case, the open source organization Open Compute ' enters this market or creates an entirely new market with innovative technology, they can 'disrupt' the status quo. Some argue that disruptive innovation creates jobs.
One of these stagnant markets is the business of selling enterprise servers. It's quite common to pay anywhere from $10,000 to $80,000 for a single standard, power-hungry x86 server.
The margins on these servers are high ' possibly in the 30%+ range. But most enterprise customers are willing to pay that premium because even at high prices, these servers make business sense. Unless, that is, you are a hyperscale organization such as Google, Facebook, or even Rackspace ' companies that buy billions of dollars worth of data centre hardware every year. For these companies, custom hardware can significantly reduce capital and operating costs (including reducing power usage) as well as generate innovative designs internally.
Early in their corporate lifetime, hyperscale organizations like Google and Facebook realized that, given their huge expenditures on data centre hardware, they could go directly to the Original Design Manufacturers (ODMs), work with them to design and build their servers, networking gear, and data centre equipment, and thus save money on lower margins. But more importantly, they could build systems that are a perfect fit for their IT business requirements.
These hyperscale companies had the freedom to start from scratch ' they could change rack sizes, power solutions, add in software defined networking ' anything they felt was important and innovative.
Open Compute is an open source extension of that internal innovative disruption, and, like Linux, it's not a purposeful attempt to disrupt to create profit. Rather, it's goal is to open up development worldwide and spur innovation through that openness.
However, despite the size of hyperscale organizations, only a small percentage of IT workers are employed at them, so a big question is: can smaller organizations get access to this hardware on a similarly smaller scale?
'Fusion sees this kind of computing spreading downmarket, from the 1,000+ unit hyperscale to the 100+ unit web scale…'
' Chris Mellor, The Register
Going farther, I can see a future (is it now?) where a systems administrator, or a developer, a hardware engineer… anyone, can use simple CAD software to build on existing, open source hardware designs, upload those designs to an ODM, and in a couple of weeks receive custom servers built to those specifications. Right now high-school students are doing the same thing with DNA, so why not with servers, or racks, networking gear, or data centre hardware?
Here are a few of the innovative ideas and hardware that came out of the 2013 Open Compute summit:
Group Hug received the most press because it allows users to easily switch out different processors, and add in new processor designs, without having to change the underlying hardware. Given that Moore's Law is still fairly accurate, there are large gains to be had if it's possible to plug a new CPU into a server without having to change out the entire server chassis. Further, if low power CPUs make sense for the workload, they can be swapped in as well.
Group Hug "shows for the first time that we can have different vendor CPU technologies on the same system," Frankovsky told attendees at an Open Compute Summit in Santa Clara, Calif.
This was my favorite project at the summit.
Calxeda took the OCP Open Vault storage system, aka Knox, and replaced its SAS card with a lower power ARM server board, turning the Knox from just-a-bunch-of-disk (JBOD) into a low power server called Project Knockout. This project could be interesting to those running object storage systems such as OpenStack Swift, or those who are interested in cold storage systems.
At the OCP summit, FusionIO announced their open source design for simple PCIe solid state storage, called ioScale, that currently goes up to 3.2TB with a price point of about $3.89/GB.
This product enables Facebook, and soon other organizations, to create small servers that have large amounts of extremely fast, very durable solid state storage, and to do so without having to rely on RAID cards and multiple SATA-based SSD drives. FusionIO suggests that the simplicity of the design will increase reliablity as compared to RAID and SSD configurations, essentially because there are fewer parts to break.
One of the more interesting presentations was by Facebook on its concept of the disaggregated rack. As I understood it, this is the concept of splitting out servers into specific 'sled' types, so that each sled provides a single main resource, such as ram or flash storage. Then the entire rack, or group of racks, really becomes the 'server.'
- 128GB to 512GB or more
- Processor: FPGA, ASIC, mobile or desktop processor
- 2 processors
- 8 or 16 DIMM slots
- No disk (network boot)
- Big NIC ' 10Gbps or more
- Essentially the Open Vault
- 500GB to 8TB of flash storage
- 600k IOPS
Facebook understands its workload so well, it can create a few different server types that match specific services. Building out the concept and design of the disaggregated rack will help it to accomplish that.