About Ravi Netravali

Biography

Ravi Netravali is an associate professor of Computer Science at Princeton University, directing the Systems and Networking Systems (SNS) Group. His research spans systems for machine learning (ML), large language model (LLM) serving, video systems, and networked systems, with contributions to top venues like PLDI, OSDI, and NSDI. [bio]

LLM Systems and Inference Optimization

Netravali's work on LLM infrastructure tackles efficiency bottlenecks in serving, inference, and agentic workflows. FailFast uses diffusion LLMs for speculative decoding, dynamically adjusting speculation lengths to balance speed and quality.[5] Aragog and its dynamic just-in-time model routing decouple configuration selection for scalable agentic workflows, improving throughput under fluctuating loads.[6][7] LessIsMore introduces training-free sparse attention via cross-head token aggregation, cutting reasoning latency by 1.13× end-to-end without accuracy loss.[8][9] SpecReason accelerates large reasoning models (LRMs) with semantic speculative reasoning on thinking tokens.[10][13] METIS optimizes retrieval-augmented generation (RAG) through adaptive scheduling and configuration.[18][19] Marconi enhances prefix caching for hybrid LLMs with reuse-aware policies.[20][21]

Video Systems and Conferencing

A major focus is real-time video, from rate control to hardware acceleration. Mowgli applies offline RL to production telemetry for rate control, boosting bitrates 15-39% over GCC without online exploration.[2][22][23] Scallop offloads SFU media operations to Tofino ASICs in an SDN-style design, scaling 210× and cutting latency 26×.[14][15] Other innovations include Dashlet for swipe-uncertainty in short video QoE,[30] MadEye for PTZ camera optimization,[26] and Legilimens for continuous learning on edge SoCs.[4][11]

Networked Systems and Hardware Offload

Netravali reframes networks as active compute resources. SmartNICs are positioned for AI inference offloading due to packet-processing alignment.[16][17] ABC enables precise congestion control on wireless links with single-bit feedback.[40][41] Application-centric design integrates app insights into networking.[24] Scallop exemplifies SDN-inspired hardware for video SFUs.[14][15]

Distributed Systems Debugging and Observability

Tools like Snicket provide query-driven tracing for microservices,[1] Lumos uses provenance-guided debugging,[3] and Revelio generates ML-assisted debugging queries.[34]

Edge and Resource-Constrained ML

Systems address memory, compute limits: GEMEL merges models for edge video analytics,[33] Apparate uses early exits for ML inference,[25] Bamboo fills pipeline bubbles for preemptible DNN training,[31] and Legilimens enables on-device continuous learning.[4][11]

Privacy, Security, and Other Systems

Privid introduces duration-based differential privacy for video analytics.[29][35] Guillotine proposes hardware-software co-design for AI containment.[12] MARVOLO augments ML malware detection,[28] Canvas isolates remote memory swaps,[32] Dorylus scales GNN training on CPUs,[37] and Gringotts incentivizes P2P video with crypto proofs.[42]

Web and Miscellaneous

JavaScript-aware crawler Java cuts web archive storage 41%.[27] Khameleon prefetching for DVE apps achieves sub-100ms latencies.[39] Boggart accelerates general video analytics with imprecise pre-filtering.[36] SAMU offloads GC tracing in disaggregated datacenters.[38]