Scaling Federated Systems at Meta: Innovations in Analytics and Learning

At Meta, we are advancing the scalability and efficiency of federated systems through innovations in both Federated Analytics (FA) and Federated Learning (FL). Our FA system is designed to facilitate privacy-preserving analytics for billions of devices, addressing the key challenges of scalability, resource efficiency, and data privacy. By leveraging one-shot algorithms, batch processing, and predictable query loads, FA achieves efficient large-scale data processing while ensuring robust privacy safeguards through Trusted Execution Environments (TEEs). The system supports flexible ad-hoc querying with rapid iteration cycles, minimizing resource consumption even on constrained devices.

Simultaneously, we have enhanced our internal FL simulation framework, FLSim, to meet the demands of large-scale distributed learning. We addressed previous scalability bottlenecks by integrating FLSim with asynchronous Remote Procedure Call (RPC) communication protocol. As a result, FLSim can now simulate FL training with a throughput of 200,000 users—each with 50 samples—with 10 million samples per minute for a small three-layer neural network over an 8x8 distributed cluster.

The synergy of FA's scalable architecture and FLSim's optimized FL capabilities have enabled Meta to deploy internal use-cases leveraging federated technologies, with a pipeline of additional applications in development.

Authors:
PPML Team: Anjul Tyagi, Othmane Marfoq, Luca Melis, Aparna Aketi
Pytorch Edge Team: Diego Palma Sánchez, Harish Srinivas