A couple of years ago, our own Paul McLellan gave us a report on the 2013 Linley Microprocessor Conference with a provocative headline: “Server Shift to ARM Becomes a Stampede”, a title right off one of the Linley slides. 64-bit ARMv8 architecture was relatively new to the game, and ARM share in networking platforms was just a sliver compared to PowerPC, X86, and MIPS.
Since that time, Freescale has become the last PowerPC company standing with AppliedMicro bolting for ARM, and LSI Axxia changing hands a couple of times (and now part of Intel). MIPS has held its ground based on a lead in network processors from Broadcom and Cavium et al. Intel is obviously still a power in servers. I dug around and didn’t find Linley’s updated data, but I suspect the ARM “server stampede” is still mostly on the outskirts of town, trampling PowerPC.
What is the holdup? As someone who has worked around the carrier-grade operating system types for 20 years, I can tell you it’s just not as easy as declaring you have a chip ready for networking infrastructure. The software gate has to be opened. At that 2013 conference, there was some violent agreement. The Linley team suggested it was too expensive for vendors or carriers to create and maintain large code bases on more than one architecture. Broadcom said their stack ran on multiple CPU architectures, which is great – if you happen to use a Broadcom chip and their software.
Let’s put aside the gory details of CPU register sets, SIMD instructions, and even endianness. All those can trip up portability in a huge code base. All can be handled if a given SoC vendor has a good relationship with an operating system vendor, and they provide the right tools for application developers. What is still missing is network-level integration and interoperability – telecom and IoT infrastructure environments are almost never greenfields. Fortunately, the drive for a new architecture and open source support for it are underway.
The real shift of importance is NFV, or network functions virtualization. Instead of building dedicated network appliances, such as load balancers, firewalls, and packet shapers, virtual machines take over the functions running on familiar servers with the necessary network interfaces. Virtualized network functions, or VNFs, are chained into services. These are then orchestrated into carrier-grade services, supporting features like network monitoring, availability, security, and billing.
Open source communities have sprung up around different pieces of the network puzzle. The Linux Foundation’s OPNFV project has established an open NFV reference platform with an emphasis on standardizing APIs between the elements. OpenDataPlane is tackling open APIs for the data plane, which harmonizes different network processor and protocol offload strategies, while maintaining concepts like QoS. There are host of other related initiatives, including OpenStack (cloud), OpenDaylight (SDN), and Open vSwitch (virtualized network switching) to name a few.
What constitutes a “server” is changing, and the cloud, mobile, and the IoT are stirring up a new round of infrastructure wars. This time, it isn’t about the data center and enterprise transactional databases and web services – it’s about moving more packets more efficiently.
ARM understands how to get software on their architecture. With Wind River paired with Intel, and MontaVista linked with Cavium, a logical choice for ARM to turn to for carrier-grade expertise is Enea. An ARM-based OPNFV reference platform, running pieces of Enea COSNOS on an AMD “Hierofalcon” Embedded R-Series SoC with eight ARM Cortex-A57 cores, was demonstrated at NFV World earlier this month. The developer release is being folded back into OPNFV, so more ARM developers can get on the trail.
Qualcomm is among the latest to join AMD, AppliedMicro, Cavium, Freescale and others in the ARM networking infrastructure SoC mix. I don’t think we should call this segment “server chips” anymore, because the network processor and data plane are a big part of the territory. This may resemble more of an orderly cattle drive than a stampede, and it may take longer to move through town than initial projections.
With the open source community behind the move to NFV, the likely move for ARM licensees – working with downstream consumers like Cisco, Facebook, and Google – is to customize SoCs for the task. The move to workload-optimized customization is exactly what sank Intel in mobile, and it could do the same in networking infrastructure if Intel doesn’t respond fast enough. (That Intel-Altera story just won’t go away, will it? “They love us, they love us not, they love us, …”)
The next big challenge for ARM will be develop an infrastructure platform channel. There is quite a gap between larger firms like Dell and HP and Oracle, and smaller companies like Advantech, Kontron, and Radisys. Several have ventured in and not been able to stay, partly because they were burned in telecom boom-bust cycles. This is what people mean when they say ARM can’t succeed without a high-volume server presence. However, PowerPC had a pretty good run in that space between, serving networking OEMs. Will ARM be next up?Share this post via: