Microsoft researchers and engineering teams will present 11 accepted papers at NSDI ’26, running May 4–6, and are participating on the program committee and in other organizational roles. The company is a returning sponsor of the conference, signaling continued engagement with the networked‑systems research community and a focus on performance and reliability improvements for datacenters and AI stacks.
Monday’s sessions (May 4) feature DroidSpeak (2:00–3:20 PM) and Eywa (3:50–5:30 PM). DroidSpeak enables LLMs that share identical architectures to partially reuse KV caches across fine‑tuned variants, producing up to 4× higher throughput and faster responses while keeping output quality impacts minimal. Eywa uses LLMs to automatically extract protocol models from natural‑language sources for model‑based testing; the authors report discovering 33 bugs in widely used network protocol implementations, 16 of which were previously unknown.
Tuesday’s program (May 5) includes Octopus (2:00–3:20 PM) and HEDGE (3:50–5:30 PM). Octopus proposes a switch‑free architecture for disaggregated CXL memory pods that reduces cost and scales to multi‑rack deployments; on a three‑server hardware prototype, Octopus RPCs ran 3.2× faster than in‑rack RDMA and 2.4× faster than CXL switches. HEDGE tackles wavelength‑specific faults in optical networks by combining link‑local and global resilience controls to preserve stable capacity and optimize traffic flow, matching existing systems’ throughput while reducing disruptions.
Wednesday’s session (May 6, 9:00 — 10:20 AM) presents AVA, an approach for open‑ended video analytics that fuses event knowledge graphs with agentic retrieval over vision‑language models. The authors also introduce AVA‑100, a benchmark made of eight videos each exceeding 10 hours and 120 manually annotated, diverse question — answer pairs; AVA attains 75.8% accuracy on this ultra‑long, open‑world evaluation. Beyond paper presentations, the conference program highlights crosscutting community activities and outreach. A video series titled "On Second Thought," hosted with Sinead Bovell and featuring company experts, is presented alongside formal conference roles to underscore the teams’ broader contributions to systems and networking research.
The papers carry concrete operational implications. KV‑cache reuse can raise serving throughput and reduce latency for families of fine‑tuned LLMs. Eywa demonstrates that LLM‑driven model extraction and model‑based testing can surface real, previously unknown protocol bugs. Octopus points to lower‑cost, scalable options for disaggregated memory in multi‑rack environments. HEDGE outlines a practical route to more resilient optical transport when link performance fluctuates. AVA and AVA‑100 provide both a benchmark and techniques for evaluating and building video‑analytics systems that operate over ultra‑long timescales.
Sources
Replies (0)
No replies in this topic yet.
