With Software Defined: Networking (SDN), Storage and Data Center movements firmly entrenched, one might believe there’s not much opportunity for innovation in dedicated hardware implemented in silicon. Several sessions at the 2014 Hot Interconnects conference, especially one from ARM Ltd, indicated that was not the case at all.
With the strong push for open networks, chips have to be much more flexible and agile, as well as more powerful, fast and functionally dense. Of course, there are well known players for specific types of silicon. For example: Broadcom for switch/routers; ARM for CPU cores (also Intel and MIPS/Imagination Technologies), many vendors for System on a Chip (SoC)- which includes 1 or more CPU core(s)-mostly from ARM (Qualcomm, Nvidia, Freescale, etc), and Network Processors (Cavium, LSI-Avago/Intel, PMC-Sierra, EZchip, Netronome, Marvell, etc), bus interconnect fabrics (Arteris, Mellanox, PLX /Avago, etc).
What’s not known is how these types of components, especially SoC’s, will evolve to support open networking and software defined networking in telecom equipment (i.e. SDN/NFV). Some suggestions were made during presentations and a panel session at this year’s excellent Hot Interconnects conference.
We summarize three invited Hot Interconnects presentations related to network silicon in this article. Our follow on Part II article will cover network hardware for SDN/NFV based on an Infonetics presentation and service provider market survey.
- Data & Control Plane Interconnect Solutions for SDN & NFV Networks, by Raghu Kondapalli, Director of Strategic Planning at LSI/Avago (Invited Talk)
Open networking, such as SDN (Software Defined Networking) and NFV (Network Function Virtualization) provides software control of many network functions. NFV enables virtualization of entire classes of network element functions such that they become modular building blocks that may be connected, or chained, together to create a variety of communication services.
Software defined and functionally disaggregated network elements rely heavily on deterministic and secure data and control plane communication within and across the network elements. In these environments scalability, reliability and performance of the whole network relies heavily on the deterministic behavior of this interconnect. Increasing network agility and lower equipment prices are causing severe disruption in the networking industy.
A key SDN/NFV implementation issue is how to disaggregate network functions in a given network element (equipment type). With such functions modularized, they could be implemented in different types of equipment along with dedicated functions (e.g. PHYs to connect to wire-line or wireless networks. The equipment designer needs to: disaggregate, virtualize, interconnect, orchestrate and manage such network functions.
“Functional coordination and control plane acceleration are the keys to successful SDN deployments,” he said. Not coincidently, the LSI/Avago Axxia multicore communication processor family (using an ARM CPU core) is being positioned for SDN and NFV acceleration, according to the company’s website. Other important points made by Raghu:
- Scale poses many challenges for state management and traffic engineering
- Traffic Management and Load Balancing are important functions
- SDN/NFV backbone network components are needed
- Disaggregated architectures will prevail.
- Circuit board interconnection (backplane) should consider the traditional passive backplane vs. an active switch fabric.
Axxia 5516 16-core communications processor was suggested as the SoC to use for a SDN/NFV backbone network interface. Functions identified included: Ethernet switching, protocol pre-processing, packet classification (QoS), traffic rate shaping, encryption, security, Precision Time Protocol (IEEE 1588) to synchronize distributed clocks, etc.
Axxia’s multi-core SoCs were said to contain various programmable function accelerators to offer a scalable data and control plane solution.
Note: Avago recently acquired semiconductor companies LSI Corp. and PLX Technology, but has now sold its Axxia Networking Business (originally from LSI which had acquired Agere in 2007 for $4 billion) to Intel for only $650 million in cash. Agere Systems (which was formerly AT&T Micro-electronics- at one time the largest captive semiconductor maker in the U.S.) had a market capitalization of about $48 billion when it was spun off from Lucent Technologies in Dec 2000.
- Applicability of Open Flow based connectivity in NFV Enabled Networks, by Srinivasa Addepalli, Fellow and Chief Software Architect, Freescale (Invited Talk)
Mr. Addepalli’s presentation addressed the performance challenges in VMMs (Virtual Machine Monitors) and the opportunities to offload VMM packet processing using SoC’s like those from Freescale (another ARM core based SoC). The VMM layer enables virtualization of networking hardware and exposes each virtual hardware element to VMs.
“Virtualization of network elements reduces operation and capital expenses and provides the ability for operators to offer new network services faster and to scale those services based on demand. Throughput, connection rate, low latency and low jitter are few important challenges in virtualization world. If not designed well, processing power requirements go up, thereby reducing the cost benefits,” according to Addepalli.
He positioned Open Flow as a communication protocol between control/offload layers, rather than the ONF’s API/protocol between the control and data planes (residing in the same or different equipment, respectively). A new role for Open Flow in VMM and vNF (Virtual Network Function) offloads was described and illustrated.
The applicability of OpenFlow to NFV1 faces two challenges, according to Mr. Addepalli
- VMM networking
- Virtual network data path to VMs/
Note 1. The ETSI NFV Industry Specification Group (ISG) is not considering the use of ONF’s Open Flow, or any other protocol, for NFV at this time. It’s work scope includes reference architectures and functional requirements, but not protocol/interface specifications. The ETSI NFV ISG will reach the end of Phase 1 by December 2014, with the publication of the remaining sixteen deliverables.
“To be successful, NFV must address performance challenges, which can best be achieved with silicon solutions,” Srinivasa concluded. [Problem with that statement is that the protocols/interfaces to be used for fully standardized NFV have not been specified by ETSI or any standards body. Hence, no one knows the exact combination of NFV functions that have to perform well]
- The Impact of ARM in Cloud and Networking Infrastructure, by Bob Monkman, Networking Segment Marketing Manager at ARM Ltd.
Bob revealed that ARM is innnovating way beyond the CPU core it’s been licensing for years. There are hardware accelerators, a cache coherent network and various types of network interconnects that have been combined into a single silicon block that is showed in the figure below:
Bob said something I thought was quite profound and dispels the notion that ARM is just a low power, core CPU cell producer: “It’s not just about a low power processor – it’s what you put around it.” As a result, ARM cores are being included in SoC vendor silicon for both networking and storage components. Those SoC companies, including LSI/Avago Axxia and Freescale (see above), can leverage their existing IP by adding their own cell designs for specialized networking hardware functions (identified at the end of this article in the Addendum).
Bob noted that the ARM ecosystem was conducive to the disruption now being experience in the IT industy with software control of so many types of equipment. The evolving network infrastructure – SDN, NFV, other Open Networking- is all about reducing total cost of ownership and enabling new services with smart and adaptable building blocks. That’s depicted in the following illustration:
Bob stated that one SoC size does not fit all. For example, one type of Soc can contain: high performance CPU, power management, premises networking, storage & I/O building blocks. While one for SDN/NFV might include: a high performance CPU, power management, I/O including wide area networking interfaces, and specialized hardware networking functions.
Monkman articulated very well what most already know: that the Networking and Server equipment are often being combined in a single box (they’re “colliding” he said). [In many cases, compute servers are running network virtualization (i.e.VMWare), acceleration, packet pre-processing, and/or control plane software (SDN model).] Flexible intelligence is required on an end-to-end basis for this to work out well. The ARM business model was said to enable innovation and differentiation, especially since the ARM CPU core has reached the 64 bit “inflection point.”
ARM is working closely with the Linaro Networking and Enterprise Groups. Linaro is a non-profit industry group creating open source software that runs on ARM CPU cores. Member companies fund Linaro and provide half of its engineering resources as assignees who work full time on Linaro projects. These assignees combined with over 100 of Linaro’s own engineers create a team of over 200 software developers.
Bob said that Linaro is creating an optimized, open-source platform software for scalable infrastructure (server, network & storage). It coordinates and multiplies members’ efforts, while accelerating product time to market (TTM). Linaro open source software enables ARM partners (licensees of ARM cores) to focus on innovation and differentiated value-add functionality in their SoC offerings.
Author’s Note: The Linaro Networking Group (LNG) is an autonomous segment focused group that is responsible for engineering development in the networking space. The current mix of LNG engineering activities includes:
- Virtualization support with considerations for real-time performance, I/O optimization, robustness and heterogeneous operating environments on multi-core SoCs.
- Real-time operations and the Linux kernel optimizations for the control and data plane
- Packet processing optimizations that maximize performance and minimize latency in data flows through the network.
- Dealing with legacy software and mixed-endian issues prevalent in the networking space
- Power Management
- Data Plane Programming API:
For more information: https://wiki.linaro.org/LNG
ODP embraces and extends existing proprietary, optimized vendor-specific hardware blocks and software libraries to provide inter-operability with minimal overhead. Initially defined by members of the Linaro Networking Group (LNG), this project is open to contributions from all individuals and companies who share an interest in promoting a standard set of APIs to be used across the full range of network processor architectures available.]
Author’s Note: There’s a similar project from Intel called DPDK or Data Plane Developer’s Kit that an audience member referenced during Q &A . We wonder if those APIs are viable alternatives or can be used in conjunction with the ONF’s OpenFlow API?
Next Generation Virtual Network Software Platforms, along with network operator benefits, are illustrated in the following graphic:
Bob Monkman’s Summary:
- Driven by total cost of ownership, the data center workload shift is leading to more optimized and diverse silicon solutions
- Network infrastructure is also well suited for the same highly integrated, optimized and scalable solutions ARM’s SoC partners understand and are positioned to deliver
- Collaborative business model supports “one size does not fit all approach,” rapid pace of innovation, choice and diversity
- Software ecosystem (e.g. Linaro open source) is developing quickly to support target markets
- ARM ecosystem is leveraging standards and open source software to accelerate deployment readiness
In a post conference email exchange, I suggested several specific networking hardware functions that might be implemented in a SoC (with 1 or more ARM CPU cores). Those include: Encryption, Packet Classification, Deep Packet Inspection, Security functions, intra-chip or inter-card interface/fabric, fault & performance monitoring, error counters?
Bob replied: “Yes, security acceleration such as SSL operations; counters of various sorts -yes; less common on the fault notification and performance monitoring. A recent example is found in the Mingoa acquisition, see: http://www.microsemi.com/company/acquisitions ”
End NOTE: Stay tuned for Part II which will cover Infonetics’ Michael Howard’s presentation on Hardware and market trends for SDN/NFV.