Taboola Makes AI Real

Personalized recommendations have changed the way brands reach their customers effectively. Taboola is the world’s largest discovery platform, delivering content recommendations to billions of consumers on many of the world’s top sites. We recently sat down with Ariel Pisetzky, Vice President of IT and Cyber, to learn how Taboola uses AI to successfully drive their business. Taboola provides the right recommendation 30 billion times daily across four billion web pages, processing up to 150,000 requests per second.

A few years ago, Mr. Pisetzky and his team required a modernized infrastructure to support Taboola’s growth and improve the experience of their customers and advertisers.

Delivering Taboola’s services requires extraordinary computing power and simplified management to attain the maximum performance to serve clients and users worldwide. The company turned to AI, because it would allow them to dynamically respond to inquiries using inferencing and deep learning capabilities. Success depended on being able to keep insights flowing with adaptable AI systems, innovative architecture and intuitive systems management.

The engine driving their AI solution consists of two components: front-end artificial intelligence (AI) for inferencing based on PowerEdge modular servers with Intel® Xeon® Scalable processors to process and deliver the real-time content recommendations. The back-end servers that host cutting-edge deep learning models are continually trained using sophisticated neural networks to infer user preferences.

By using PowerEdge modular servers, the IT team at Taboola can meet rapidly changing demands and enjoy the versatility and simplicity necessary to support a building block approach. The team is able to cost-effectively use the same servers interchangeably as AI inferencing nodes, database servers or storage nodes with very simple configuration changes. Each request coming into a front-end data center runs the AI-driven inferencing algorithms in a unique, ultra-fast process that delivers a relevant recommendation within 50 milliseconds.

Taboola took full advantage of the built-in performance acceleration of 2nd Gen Intel Xeon Scalable processors—together with the highly optimized Intel Math Kernel Library for Deep Neural Networking (Intel MKL-DNN). Taboola was able to initially enhance its performance by a factor of 2.5x or more with their modernized infrastructure. Then, gaining the efficiencies of Kubernetes within the software layer—including the operating system, TCP/IP stack, load balancing and more— Mr. Pisetzky’s team went much further.

“With PowerEdge servers and Intel Xeon Scalable processors, we now get up to six times the performance on our AI-based inferencing compared to when we started,” states Pisetzky. “This helps reduce our costs, and we believe there’s a lot more to be gained over time.”

For the back-end data centers running deep learning-based models to accurately and reliably train the Taboola models, the Dell EMC PowerEdge R740xd servers with their lightning-fast accelerators were the answer.

“Training is much different from the real-time inferencing we do on the front end. The demands aren’t in terms of response times, but rather the time it takes to process large volumes of data. PowerEdge R740xd servers provide the performance to access our massive data to train our models and push them back to our front-end data center for inferencing. We’re using Vertica, Cassandra and MySQL databases across a variety of nodes,” states Mr. Pisetsky.

Today, the company takes a more holistic view of its data centers as high-performance computing (HPC) clusters, which are able to process an enormous number of requests per second. Rather than just add servers or racks, Taboola looks at everything as a single HPC machine, and reshuffles servers to achieve significant performance improvements and greater cost efficiencies.

The next step in building Taboola’s solution was determining the most efficient and cost effective way to manage this large global footprint with a small IT team of 12 Site Reliability Engineers across nine global data centers. The team turned to iDRAC, which allows them to deploy servers with the touch of a button. They can easily update servers across their data centers and ensure the BIOS and firmware settings are identical across all servers.

The results Taboola has delivered to their users are amazing. Today, different people can go to the same page and receive personalized recommendations relevant to them, all without Taboola knowing who you are. AI has provided Taboola with the ability to take their business to the next level with impressive results. They can now provide personalized services, better user experiences and better results for their end users, advertisers and publishers.