It was interesting to see some in-depth discussions of Software Defined Networking (SDN) at last week’s Linley Tech Processor Conference in San Jose. Typically, these conferences are focused on processors and related chips such as memory, but this time there was a significant amount of time dedicated to software topics. Both SDN and OpenFlow generated some interesting conversations.
During a session titled “Meeting the Networking Challenges Posed by Cloud Computing and SDN”, Eliot Rosen, Product Line Director at Broadcom, discussed some of the key trends in data centers and their implications for new processor architectures.
Rosen began his presentation by discussing the different challenges associated with four major data center applications: application servers, video servers, web/e-commerce and financial trading. He explained that, despite the differences between these applications, they are each expected to require a 10x-40x speed improvement, 10x higher port density and 10x lower latency over the next decade.
As Rosen discussed, multicore processors for data centers need to evolve to address an exponential increase both in user data and also in associated unstructured data. This involves the centralization of hardware services to reduce the overall data center footprint and power, as well as the increasing adoption of virtualization with intelligent virtualization layers and the growth in number of VMs caused by server offload and intelligent network solutions. He reviewed the requirement for mesh connectivity in next-generation data centers, driven by more clustered applications, more East-West traffic, more VMs and increasing VM migration.
Rosen summarized the basic principles of SDN, in terms of the separation between the networking control plane and data plane, thereby centralizing control plane tasks and simplifying network organization. He pointed to two key drivers for SDN. First, SDN-based initiatives such as OpenFlow are driven by carriers in response to real-world requirements, recognizing that current network architectures are too costly, too inflexible and ultimately no longer sustainable. Second, SDN brings simplicity to an overly complex world, because given the impact of moving applications to the cloud and the requirement to introduce policy control for application access and security, networks will increase in complexity on many levels.
Finally, Rosen presented a detailed overview of how the Broadcom XLP multicore processor has been designed and optimized for SDN workloads. The control plane and services are overlaid on the XLP, so that the control plane stack operates on one or more cores, while incremental services and/or next-generation capabilities operate run on separate cores and on-chip hardware accelerators. He showed an in-depth example of a complete OpenFlow-based software architecture for the XLP. He illustrated the advantages of the XLP for SDN in terms of key features such as the dual-issue superscalar processor core, out-of-order execution, the tri-level cache, quad threading, the advanced fast messaging network, autonomous accelerators, integrated grammar engine and multi-socket coherence.
At 6WIND, we are very familiar with the details of the XLP processor architecture. Our experts worked closely with Broadcom’s engineers as we optimized the 6WINDGate™ packet processing software to best exploit the various performance features of the platform. As we described when we announced our support for the XLP, the 6WINDGate software is optimized to take full advantage of the XLP processor’s advanced Fast Messaging Network™, fully-autonomous programmable processing engines and tri-level cache architecture to achieve up to 40Gbps of IPsec and 60 million packets per second of IP Forwarding performance on a single XLP832 processor running at 2GHz while ensuring that a subset of the cores remain available for application software.
Were you at the Linley Tech Processor Conference last week? What were some of the key messages and trends that you took away from the conference?