There is a shift happening in the world of Artificial Intelligence requiring a new breed of servers, storage and cloud networks. Artificial intelligence applications for patterns, photos and speech recognition have driven a processor evolution from CPUs to NPUs to now, GPUs. Networking is witnessing a parallel evolution and pushing the scale of shuttling massive data between machines. It creates an ever-increasing need for control over the way networks are rebuilt. Building these networks requires both, programmable paths to drive intelligence and uncompromised performance. Doing both hasn’t been easy until now.
New Barefoot Tofino
Barefoot Networks is Arista’s new merchant silicon partner with the “Tofino” family achieving both vectors without compromise. It redefines forwarding-plane programmability. Network engineers and operators can now control deep packet processing at the granular chip level to customize their own features. Tofino can also work with open-source P4 Language to instruct the chip how to process packets creating powerful systems. At Arista, programmability is not new and we have been working with Broadcom, Cavium and Intel silicon to achieve this. What is fascinating about Barefoot’s silicon is the stateful network awareness. The integration into the data plane via match-action pipeline updates, and lookup tables in parallel memory, overcomes many operational issues such as prototyping protocols, rapid deployment and network telemetry.
Introducing Arista 7170 Multifunction Platform
The Arista 7170 Series is bringing a new generation of programmable packet processors for 64 ports of 100G in a single fixed leaf. Traditionally, building a chip that was both programmable and extremely fast wasn't doable without tradeoffs but Arista’s 7170 defies tradition. It integrates a suite of capabilities that previously ran on a host, VM or NIC into a single one Rack-Unit (RU) switch. Examples of advanced features include tunnel termination, traffic filtering, Network Address Translation (NAT) and deep packet inspection. Additionally, network operations that required routers or costly networking devices can now be performed in the 7170 pipeline. The Arista Networks 7170, available now, leverages this programmable capability with Arista EOS’s comprehensive L2 and L3 features, as well as added value features for higher scale. Examples include - tunnel scaling in multi-tenant datacenters, applying network security segmentation at high scale and real-time network telemetry for time-stamping, visibility and packet capture.
Cognitive Cloud Networking Era
At Arista Networks, programmability is not new. Our platforms have been programmable for many years across control, data and management planes, leveraging Broadcom, Cavium and Intel silicon. What is interesting in the Arista 7170, is the ability to offer more network awareness and change the behavior without costly replacements or upgrades. The advanced Tofino silicon from Barefoot changes legacy fixed designs to flexible pipelines with faster time to deployment, ushering in a key step to modern networking.
Arista’s state of the art EOS with programmable data planes takes the network industry well beyond standard data movement to cognitive controls. Our sophisticated customers for peering and cloud exchange and enterprises with diverse applications welcome this next frontier. Arista Networks, along with advances in modern silicon, is helping to feed the insatiable appetite of this new age of cognitive cloud networking.
Jayshree Ullal - 5 June 2018