Mobile operators which embarked early on the process of virtualizing their networks may wish they had waited. Early deployment of Network Functions Virtualization (NFV) relied on first generation technology, such as OpenStack cloud infrastructure and virtual machines (VMs), which is now seen as cumbersome and outdated. The next push will be to migrate to highly flexible, disaggregated technologies like containers and microservices, in order to support fully cloud-native core networks and, eventually, RANs.
The pain of having to update a hard-won virtualized network at such an early stage is not unique to operators. Even the webscale companies, which defined so many of the rules for the cloud-based telecoms era, have felt similar pain. Amazon’s CTO, Werner Vogels, told the AWS re:Invent conference last week in Las Vegas that one of his biggest tasks had been a multiyear effort to change the approach to virtualization, in order to support maximum efficiency and flexibility as enabled by cloud-native technology.
“Virtualization has been the bread and butter of the compute parts of any cloud environment from day one [including in 1960s mainframes],” he said. “It’s one of those major technical underpinnings that has really allowed cloud computing to become as big as it is.” But the old monolithic approach to virtualization was cumbersome, and saw guest services competing for the same network resources, slowing performance and adding latency. This has led AWS to “push the boundaries of virtualization,” most recently with its Nitro platform, which separates the underlying hardware with a light hypervisor.
Key steps included offload of network components, and other specialized processing, to multiple cards sitting on the same server, leaving bulk tasks on the CPU. Network management, monitoring, security and storage are moved off the host platform as microservices, resulting in a lighter hypervisor and far greater flexibility and performance. This is a key enabler of the Outposts on-premises edge cloud offering (see separate item).
Nitro embeds networking functions in a separate card, and was completed last year when Amazon removed “all the other pieces of the hypervisor”, shifting these functions into the control plane on a separate card. That card, the C5, comes close to bare metal in performance, Vogels claimed, because the hypervisor is so thin.
Nitro controls all Amazon’s compute environments today and future development will bring performance even nearer to bare metal, while creating “a world where we have all these devices and they have an API”, even if that API is a hardware item like a PCI bus.
Underlying the increasingly high performance cloud environment is new chip technology. Like Google and Microsoft, AWS has invested rising amounts of money and resource in designing its own processors, to push performance and power efficiency to its limits (as a key element in competitive cost models), while reducing dependence on Intel or even ARM-based processor suppliers.
AWS has harnessed the ARM core itself to design the Graviton2 platform, which was shown off at the re:Invent event.
AWS CEO Andrew Jassy claimed the new chip has 40% better price/performance than the latest generation of Intel-x86 processors. These reported gains are based on ARM’s
64-bit Neoverse N1 platform, which it unveiled a year ago. This was a clear bid for a bigger place in cloud processors for ARM architectures, and is optimized for 7-nanometer processes and for compute speed.
Working with its Israel-based Annapurna Labs unit, AWS entered a close cooperation with ARM on the N1 IP, resulting in a fully working chip less than a year after the N1 launch. This shows a webscaler bringing the speed and agility for which the cloud giants are known into the silicon area, where long lead times are the norm. Chip analyst Patrick Moorehead said: “This is about as agile a silicon development lifecycle as one can imagine.”
While this may dismay conventional processor suppliers, including Intel and ARM supporters, it will be a hugely valuable cooperation for ARM itself, proving that its architecture can be optimized for the highest performance applications, and that it has the potential to extend its reach in the hyperscale environment – the biggest source of demand for performance processor chips.
“While lots of companies including ourselves have worked with x86 processors for a long time … if we wanted to push the price/performance envelope it meant we had to do some innovating ourselves,” Jassy commented.
Graviton2 will initially power instance based on the Amazon Elastic Compute Cloud (EC2), including its Elastic MapReduce (EMR) big data platform, Elastic Load Balancing, ElastiCache and other AWS services. All these are built on the Nitro System, which handles offload and acceleration of demanding workloads, and cloud-native virtualization of compute, storage and network services.
The first test will be to make AWS’s internal systems more efficient, reducing its costs and delivering an enhanced experience for generic services. The next will be to translate the new efficiencies and performance to specific customer workloads, improving its competitive edge against Azure and Google in critical areas such as machine learning and, for operators, virtualized RAN. Of course, AWS is not alone in developing its own chip technologies – Google and Alibaba, in particular, are active in this field, especially targeting optimized processing for artificial intelligence (AI).
ARM will hope that Graviton2 provides a strong credibility boost for its Neoverse architecture, which was developed at great expense to target the webscale platforms, and could now lure other cloud providers to the system, either directly or through their semiconductor suppliers.