The latest instalment
of the annual Open Compute Project (OCP) Summit, which was held March 8-9 in Silicon Valley, brought new open source designs for next-generation data
centres. It is six years since Facebook launched OCP and it has grown into
quite an institution. Membership in the group has doubled over the past year to
195 companies and it is clear that OCP is having an impact in adjacent sectors
such as enterprise storage and telecom infrastructure gear.
The OCP was never intended to be a
traditional standards organisation, serving more as a public forum in which
Facebook, Microsoft and potentially other big buyers of data centre equipment
can share their engineering designs with the industry. The hyper-scale cloud market,
which also includes Amazon Web Services, Google, Alibaba and potentially others
such as IBM and Tencent, are where the growth is at. IDC, in its Worldwide
Quarterly Cloud IT Infrastructure Tracker, estimates total spending on IT
infrastructure products (server, enterprise storage and Ethernet switches) for
deployment in cloud environments will increase by 18% in 2017 to reach $44.2
billion. Of this, IDC estimates that 61% of spending will be by public cloud
data centres, while off-premises private cloud environments constitute 15% of
spending.
It is clear from previous disclosures
that all Facebook data centres have adopted the OCP architecture, including its
primary facilities in Prineville (Oregon), Forest City (North Carolina),
Altoona (Iowa) and Luleå (Sweden). Meanwhile, the newest Facebook data centres,
under construction in Fort Worth (Texas) and Clonee, Ireland are pushing OCP
boundaries even further in terms of energy efficiency.
Facebook’s ambitions famously
extend to connecting all people on the planet and it has already passed the
billion monthly user milestone for both its mobile and web platforms. The
latest metrics indicate that Facebook is delivering 100 million hours of video
content every day to its users; 95+ million photos and videos are shared on
Instagram on a daily basis; and 400 million people now use Messenger for voice
and video chat on a routine basis.
At this year’s OCP Summit, Facebook
is rolling out refreshed designs for all of its ‘vanity-free’ servers, each
optimised for a particular workload type, and Facebook engineers can choose to
run their applications on any of the supported server types. Highlights of the
new designs include:
·
Bryce Canyon, a very
high-density storage server for photos and videos that features a 20% higher
hard disk drive density and a 4x increase in compute capability over its
predecessor, Honey Badger.
·
Yosemite v2, a compute
server that features ‘hot’ service, meaning servers do not need to be powered
down when the sled is pulled out of the chassis in order for components to be
serviced.
·
Tioga Pass, a compute
server with dual-socket motherboards and more IO bandwidth (i.e. more bandwidth
to flash, network cards and GPUs) than its predecessor, Leopard, enabling
larger memory configurations and faster compute time.
·
Big Basin, a server
designed for artificial intelligence (AI) and machine learning, optimised for
image processing and training neural networks. Compared to its predecessor, Big
Basin can train machine learning models that are 30% larger due to greater
arithmetical throughput and by implementing more memory (12 to 16 Gbytes).
Facebook currently has web server
capacity to deliver 7.5 quadrillion instructions per second and its 10-year
roadmap for data centre infrastructure, also highlighted at the OCP Summit,
predicts that AI and machine learning will be applied to a wide range of
applications hosted on the Facebook platform. Photos and videos uploaded to any
of the Facebook services will routinely go through machine-based image
recognition and to handle this load Facebook is pursuing additional OCP designs
that bring fast storage capabilities closer to its compute resources. It will
leverage silicon photonics to provide fast connectivity between resources
inside its hyper-scale data centres and new open source models designed to
speed innovation in both hardware and software.
