NDM-TCP v2.0: Bringing Neural Network Intelligence to High-Speed Networks

Project repository: github.com/hejhdiss/lkm-ndm-tcp-v2 The Problem with Smart TCP Everyone wants "smarter" network protocols. Machine learning promises to optimize congestion control by learning from network patterns in ways traditional algorithms can't. The problem? Most ML-based TCP implementations are resource hogs. A typical deep learning congestion control system might consume hundreds of megabytes of RAM and require TensorFlow or PyTorch running in userspace. When we released NDM-TCP v1.0, we solved the memory problem—just 70 bytes per connection—but the CPU overhead was still too high for ultra-fast networks. At 100 Gbps, v1.0 would consume 145% of a CPU core. Clearly unacceptable. Version 2.0 fixes this. Through five carefully designed optimizations, we've reduced CPU usage from 145% to just 4-6% of a core at 100 Gbps—a 24× improvement—while maintaining the intelligence that makes NDM-TCP effective. The Five Optimizations That Made It Possible Adaptive computation is the first