Key Takeaways:

• Microsoft was everywhere- presenting at the Wednesday morning keynotes, announcing partnerships and new initiatives during Thursday engineering workshops.  Progress on the SONIC protocol stack (announced last year) and the new Project Olympus were described in detail.
• Facebook revealed details of it’s “server refresh” and 100GE fiber optic switch-to-switch Data Center network (see section below).
• Yahoo – Japan explained that lower electricity cost was the reason the company has a Data Center in Washington state.  They showed a video demonstrating how simple and quick it was to swap out an OCP hardware module in their Washington Data Center.
• Nokia talked about its vision for Telco Cloud, which is actually quite mature. In a different session, they discussed the company’s experience with OCP hardware, including backgrounds and incentives, what choices were made and why. Hardware management interfaces and related improvement needs were disclosed.
• SK Telecom shared its experiences in a trial deployment and maintenance of OCP gear, while also stating its future directions. They conducted an OCP trial for telco infrastructure with three objectives in mind: (1) Analysis of OCP System Deployment Environment, (2) Evaluation of Efficiency, Serviceability, and Performance, (3) Application Test.  The company says that the OCP ecosystem in Korea is not mature and has a relatively small market. To minimize risk, SK Telecom would very much like to collaborate with other OCP telco members on practical issues like: vendor selection, existing data center deployments, technical support, compatibility, etc.
• AT&T provided status and updates on several specifications AT&T has or will share with OCP. These include the XGS-PON specs as well as the new G.Fast specs.  [Both specs are part of Project CORD which is hosted by the Linux Foundation.   Much of the open source code done for Cord is done at ON.Lab.]  It was interesting to learn that G.Fast will operate on both low grade copper twisted pairs as well as coax.
• AT&T also revealed its future direction in disaggregated Line Termination Equipment (LTE): remove the processor from the box and replace it with redundant communications links (for failover) to a compute server.  That would enable one compute server to control and manage many LTEs.  AT&T favors dis-aggregation of network elements, which permits cloud software strategies to be used in place of firmware type applications in telecom equipment. By creating more flexible relationships between hardware and software, new availability and management models are possible.
• EdgeCore highlighted the achievements made in the OCP networking group over the last four years, overview of that latest advancements in Open Networking, and a glimpse into the future of Open Networking.  The company is one of two 100GE CWDM-4 optical switch suppliers for Facebook’s 100G Data Center network, as described in the section below.
• HPE talked about the challenges and risks of moving to an open source environment.  Citing Gartner Group market research, the company claimed it was the #1 brand for compute and storage servers and #2 in networking.
• Intel presented Rack Scale Design (Intel® RSD) – a logical architecture whose open standard APIs offer infrastructure manageability that is compatible with the hardware designs being contributed and adopted by OCP members.
• Radisys described the OCP Carrier Grade OpenRack-19 specification for the telco data center.  It’s a scalable carrier-grade rack level system that integrates high performance compute, storage and networking in a standard rack.  Key points covered included:

(1) The telecom industry’s transformation from the traditional Central Office (CO) to virtualized Data Centers and how this transition is driving demand for open hardware solutions from OCP.

(2) Why adapting the OCP base model to the telecom environment must take into account new requirements to accommodate existing Central Offices.

(3) The specific requirements for carrier-grade OCP, including Physical, Content/Workload, Management, and Networking/Interconnect.

(4) How the CG-OpenRack-19 specification aligns with the principles of OCP. Panelists will discuss in-depth how the spec meets requirements for efficient design, scale-out ready architecture, open source and creating a meaningful positive impact.

(5) A real-world example of a CG-OpenRack-19 based hardware solution being deployed in a live network.

Verizon said they will use the OCP Carrier Grade-OpenRack-19 specification to simplify data center deployments and speed up failed equipment replacement time.

• Ericsson discussed their collaboration in open source software projects in order to provide a common management agent framework and sample agent for Ericsson’s hyperscale data center solutions. In a different session, the company discussed solutions for hardware that needs to remain operational in telco data centers which may be deployed in high-temperature (up to 50 degrees centigrade = 122 degrees fahrenheit) situations. A few examples based on Ericsson’s BSP 8100 telco-grade blade servers were presented.
• Google was present, but kept a very low profile, with no keynote presentation(s) or exhibit hall booth.  OCP Operations Director & Community Manager Amber Graner wrote in an email to this author:

“Google was there. They were presenting in the Engineering Workshops for the 48v rack standard as part of the Rack and Power Group. They engaged and involved in the community.”

Personally, the best session for me was saved for end of the conference.  The very last session on the last day (Thursday):

100G Optics Deployed in Facebook’s Data Centers

Abstract:  “This engineering workshop will describe Facebook’s deployment of 100G optics in switching networks with examples of recent data center use-cases. The workshop will provide details on the optical transceiver specification and the type of fiber installed. Examples will be given of production testing used to ensure these hyper-scale networks deploy successfully, and the focus on quality that ensures they keep running. Details of the ecosystem of optical transceivers and switching equipment that supports 100G deployments will be shared.”

……………………………………………………………………………………………

At this session, a panel of Facebook employees disclosed details of their 100G SMF switch to switch deployment within all their data centers (DCs).  That’s quite an impressive accomplishment considering that most cloud DCs use 40G over MMF for switch interconnections.   

Facebook’s 100G optical interconnect spec is called CWDM4-OCP.

Composition of a typical DC Optical Link is shown below:

Facebook also named its suppliers of 100G optical transceiver modules (Finisar & ColorChip) and 100G optical Ethernet switches (EdgeCore and a Chinese /Taiwanese based ODM) that have implemented the 100G CWDM4 OCP specification.

In answer to an audience member question, Facebook said that they use traditional copper cable at various speeds (10G, 25G, etc) for interconnecting a TOR (Top of Rack) switch to a compute server.  In the future, they might use some form of fiber optics if there was a need for faster bit rate interconnections.  

When asked about migration to 400G by another audience questioner, Facebook panel chair Katharine Schmidtke said:   “While the move from 40G to 100G (within Facebook DCs) was fairly quick, 100G to 400G would likely take longer.  It’s time to catch our breath.”

Indeed, the IEEE 802.3bs 400GE standard is not expected to be completed till December 2017 or January 2018.  The timeline is here.

…………………………………………………………………………………………

References:

https://code.facebook.com/posts/1633153936991442/designing-100g-optical-connections/

2016 OCP Summit Solidifies Mega Trend to Open Hardware & Software in Mega Data Centers – Technology Blog