Telemetry data is coming into its own as a tool for network owners and application developers to better understand network performance. This information can be used by the switches in the network to reduce network congestion and minimize latency.
Telemetry can be thought of as a track coach with a stopwatch at each network element, measuring the time it takes each packet to sprint across the finish line. The coach also gathers and reports this data for every packet traversing the network.
One of the key features of the Intel® Tofino™ P4-programmable switch is in-band network telemetry (INT)—an ability to collect telemetry data on every packet in the network in real time. INT seeks out important data including: What path did a packet take? Which rules did the packet follow? How long did the packet queue at the last switch? What other packets shared the queue? When the answers to these questions are processed by network analytics software, network managers can better understand their network data flows and fix congestion issues in real time. For more information on the INT framework, here is a recent Intel editorial describing it in more detail.
While in-band telemetry is not yet ubiquitous, it has found initial traction in high performance networks with specialized applications and in choke points of large networks. Customers that can benefit from INT today include cloud and communications service providers. Intel and Arista Networks have been working with several high-performance web cluster customers to use INT to improve the granularity of instrumentation for user/application monitoring and capacity planning.
Arista Networks has been a pioneer in P4-programmable data center switching using Intel Tofino switch ASICs. The company’s Arista 7170 Series are INT-capable, 100GbE switching systems designed for big networks handling a lot of data—up to 6.4 Tbps per system. Support for INT is part of Arista's vision of network-wide visibility—a goal to provide extensive, real-time telemetry of both traffic flows and device state to network administrators.
Arista’s Extensible Operating System (EOS®) provides broad support for real-time streaming of system state and resource utilization, as well as industry standard flow monitoring capabilities such as sFlow and IPFIX. Complementing these with INT enables an additional layer of fine-grained information about the traffic flows within each individual switch. When implemented across multiple network hops, a detailed picture of real-time utilization and behavior can be built.
The programmability of the Tofino chip is an advantage in the evolving world of in-band network telemetry. Since it is based on the P4 language, unique functionality that normally requires dedicated hardware can now be enabled within the switch at line rate. As new use cases develop, the mode of operation can also be tuned programmatically without waiting for a new hardware implementation. These are critical benefits for customers looking to scale their architectures, while keeping control of costs.
Maintaining low latency network performance in the data center is more critical than ever. Arista understands this and is providing its customers with a complete toolbox of telemetry-based management solutions including INT on its Intel Tofino-based switches. Working together, Intel and Arista will help make INT capabilities available to more companies that need the technology to fine-tune network application performance and eliminate congestion.
Edward (Ed) V. Doe is vice president of the Connectivity Group and general manager of the Barefoot Division (BXD) at Intel Corporation. Doe holds a bachelor’s degree in electrical engineering from the University of Western Ontario, London, Ontario, Canada and two years of graduate studies towards a master’s degree in electrical engineering at Stanford University.