fabric_ext: One eBPF policy spans GPU, DPU, and CXL switches

Yiwei Yang and Andi Quinn's new paper introduces fabric_ext, an eBPF-based middleware compiler and runtime designed for extensible operating system policies over GPU-CXL fabrics. The system allows a single policy program to execute coherently across vastly different execution contexts: GPU hooks, driver/runtime hooks, DPU/NIC hooks, and CXL switch or near-memory hooks. The core abstraction is a "semantic movement graph" where edges describe key attributes of data movement—bytes, stride, reuse distance, read/write ratio, source/destination, ordering, alias sets, ownership—and transformations like Move, Quantize, Compress, Checksum, Filter, Reduce, Scatter/Gather, Replicate, and Persist. The fabric_ext compiler lowers this graph into per-device eBPF programs, verifier obligations, consistency-classed BPF maps, and artifacts for bpftime and dputime, enabling coherent policy enforcement across heterogeneous accelerators and network elements.

At the fabric edge, fabric_ext treats a near-Type-2 small core as a hardware-JIT and state manager: it specializes verified movement descriptors into local copy, placement, ordering, and transformation commands while the surrounding Von Neumann island of memory, DMA, and compute engines executes the dataflow. Because this dataflow is data-driven, fabric_ext also places observation logic beside the island, where queues, DMA completions, memory placement, and ownership transitions are visible as they happen. The canonical stress case is LLM prefill: attention streams KV blocks and reductions while FFN streams activations, weights, and compressible intermediates, forcing one request to cross GPU tensor execution, DPU/NIC event execution, and CXL or switch-local dataflow islands. This work demonstrates a practical path toward programmable, cross-layer data movement policies in modern AI infrastructure.