Telefonica: “I believe one of many killer apps of community programmability might be ML”
Earlier than diving into the function of a programmable information aircraft in supporting telco AI purposes, Telefonica Analysis’s Senior Analysis Scientist Eduard Marin Fabregas supplied a little bit of a historical past lesson to attendees of the Telco AI Discussion board 2.0, obtainable on demand right here. First issues first, he famous the everyday community composition of a administration aircraft to watch and configure units remotely, a management aircraft, together with protocols to populate forwarding tables, and an information aircraft made up of all of the units used to ahead packets.
This setup, Fabregas stated, was designed “with resiliency in thoughts, which was nice within the early days. However it additionally got here with some issues. One of many issues…is that the info aircraft and the management aircraft are tightly coupled collectively on the save gadget.” {Hardware}-centric routers lacked programmability, supported a set set of functionalities, and the ASICs they’re constructed with led to price and complexity. “This prevents us from innovating,” he stated. “And, after all, this has had some affect on many purposes that we might develop within the networks. AI is certainly one of them.”
The rise of software-defined (SDN) controllers, and different applied sciences like community capabilities virtualization, container administration and eBPF, have enhanced the flexibility to watch packets and information flows, and different community telemetry, and have in any other case elevated programmability. Particular to SDN, Fabregas stated, open, commonplace interfaces and disaggregation have enabled a bigger ecosystem to ship extra innovation. “And this has modified loads the best way we are able to do ML within the community.”
He continued: “Now what we are able to do is we are able to do characteristic assortment on the info aircraft, so on the router itself.” Nonetheless, inference can’t be accomplished within the routers. As an alternative, operators configure the info aircraft to pattern particular information factors usually both periodically or as triggered by outlined occasions; that data can then be forwarded to the management aircraft. “The issue is we are able to’t actually do per packet inference at reside velocity…Nonetheless this isn’t very best. However this isn’t the tip of the story.”
He stated that utilizing programmable information planes for ML duties is “one of many greatest improvements that we’ve had in a few years.” Utilizing the P4 programming language, CSPs can basically set routers as much as conduct specified operations at Terabit velocity for real-time determination making and improved community visibility. “We are able to resolve issues on the community in actual time on the routers themselves…We are able to extract many extra insights from the community.” As a result of routing and switching infrastructure sits between consumer units and the core community, “It may possibly play an vital function in lots of extra of the purposeful elements behind ML,” Fabregast stated.
Fabregast went on to explain a protocol-independent change structure (PISA) consisting of a parser, programmable pipeline and de-parser (see above picture). Primarily, he stated, PISA permits for a number of levels the place ML inference might be utilized, albeit with challenges round reminiscence. “These units supply big alternatives for parallelization, so operations that don’t depend upon one another could possibly be positioned on the identical stage.” Including ML inferencing within the information aircraft permits for extra complicated characteristic extraction and customization.
Fabregast gave the instance of anomaly detection historically accomplished within the router by sampling packets, and probably forwarding these packets to a different equipment operating an ML mannequin to resolve, for example, if a packet is malicious or not, then sending it again to the router the place a specific coverage might be utilized. Placing that ML mannequin immediately on the change might assist anomaly detection of each single packet with out affecting throughput. He described this use of a programmable information aircraft as “a primary line of protection. You might consider this as deploying a comparatively light-weight, easy ML mannequin maybe simply to detect assaults.”
Massive image, Fabregast stated, “The longer term is self-driven autonomous networks—networks that can make selections primarily based on information.” Adopting closed-loop automation for monitoring, evaluation and motion will result in adaptive, resilient networks. Developments in {hardware}, unified requirements and APIs to share data between switches, and additional mannequin improvement utilizing artificial and augmented information, will drive additional programmability. “I believe one of many killer apps of community programmability might be ML,” he stated.